KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #cli #open-source #ai #claude-code #codex #developer-tools #dsh-plugin #ai-agent #mcp #deepseek-harness

UHD

★ 55 Open GitHub ↗

Ultra-lightweight human detection. The number of parameters does not correlate to inference speed. For limited use cases, an input image resolution of 64x64 is sufficient. High-level object detection architectures such as YOLO are overkill.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# UHD
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.17790207.svg)](https://doi.org/10.5281/zenodo.17790207) ![GitHub License](https://img.shields.io/github/license/pinto0309/uhd) [![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/PINTO0309/uhd)

**Please note that the dataset used to train this model is a custom-created, ultra-high-quality dataset derived from MS-COCO. Therefore, a simple comparison with the Val mAP values of other object detection models is completely meaningless. In particular, please note that the mAP values of other MS-COCO-based models are unnecessarily high and do not accurately assess actual performance.**

- PC

https://github.com/user-attachments/assets/afca301a-9fe6-4ecd-af01-6bacbfa88e52

- ESP32-S3 - UltraTinyOD-Anc8-W32-64x64

model:https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.tar.gz

demo app:https://github.com/PINTO0309/esp-who/tree/custom/examples/ultra_lightweight_human_detection

https://github.com/user-attachments/assets/cbefde30-6d73-4a9a-8b45-1a92826e76c8

- UltraTinyOD-Anc8-W32-64x64

model:https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.tar.gz

demo app:https://github.com/PINTO0309/esp-who/tree/custom/examples/ultra_lightweight_human_detection

|Input/Output|Input/Output|Input/Output|Input/Output|Input/Output|
 |:-:|:-:|:-:|:-:|:-:|
 |<img width="64" height="64" alt="00_dist_000000075375" src="https://github.com/user-attachments/assets/e4a06595-9e75-4f75-829b-827789be75f5" />|<img width="64" height="64" alt="01_000000073639" src="https://github.com/user-attachments/assets/7f286e88-8b3d-431d-aea2-0949817fabc4" />|<img width="64" height="64" alt="02_000000051704" src="https://github.com/user-attachments/assets/847e2211-8a26-432e-b6e2-4f9b473a3bd5" />|<img width="64" height="64" alt="03_frame6_000286" src="https://github.com/user-attachments/assets/34f566d3-8d21-4a35-aa9d-1186f213ebbd" />|<img width="64" height="64" alt="08_dist_000000039322" src="https://github.com/user-attachments/assets/6d0bd140-6e20-430b-9e79-d590d3ddbe3e" />|
 |<img width="64" height="64" alt="04_000000014502" src="https://github.com/user-attachments/assets/761ce21b-cd30-4bdf-923f-fe1c89a8593c" />|<img width="64" height="64" alt="05_000000086549" src="https://github.com/user-attachments/assets/a01aac9e-374a-4daf-a0e9-990828fb93db" />|<img width="64" height="64" alt="06_frameA_000088" src="https://github.com/user-attachments/assets/5b7e7d80-9e24-43ad-b63d-ec072affdf49" />|<img width="64" height="64" alt="07_000000016314" src="https://github.com/user-attachments/assets/47686e20-3c76-40d7-9dd7-b7d5afd62500" />|<img width="64" height="64" alt="09_dist_000000074177" src="https://github.com/user-attachments/assets/3da5a712-62fd-450f-936d-32e7123621a7" />|

- Variant-S / `w ESE + IoU-aware + ReLU`

|Input 64x64|Output|Input 64x64|Output|
 |:-:|:-:|:-:|:-:|
 |<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/59b941a4-86eb-41b9-9c0a-c367e6718b4c" />|<img width="640" height="427" alt="00_dist_000000075375" src="https://github.com/user-attachments/assets/5cb2e332-86eb-4933-bf80-3f500173306f" />|<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/21791bff-2cf4-4a63-87b5-f6da52bb4964" />|<img width="480" height="360" alt="01_000000073639" src="https://github.com/user-attachments/assets/ed0f9b67-7a05-4e31-8f43-8e82a468ad38" />|
 |<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/6e5f1203-8aad-4d2a-8439-27a5c8d0a2a7" />|<img width="480" height="360" alt="02_000000051704" src="https://github.com/user-attachments/assets/20fba8fc-5bf6-4485-9165-ce06a1504644" />|<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/5e7cd0e1-8840-4a4e-b932-b68dc71f3721" />|<img width="480" height="360" alt="07_000000016314" src="https://github.com/user-attachments/assets/b9aced38-c4eb-4ee7-a6c5-294090748eca" />|
 |<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/d4b5fc1c-324b-4ed3-9ab7-35c1d022b2d0" />|<img width="480" height="360" alt="28_000000044437" src="https://github.com/user-attachments/assets/499469d3-41a3-406b-a075-8b055297bf6e" />|<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/9be24b81-0c6e-48c7-bda1-1372b6ae391f" />|<img width="480" height="360" alt="42_000000006864" src="https://github.com/user-attachments/assets/fed81463-3cd6-4d69-8fc3-5bf114d92ab8" />|
 |<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/05280dce-4faf-46e1-817c-98e4df4a9d4c" />|<img width="383" height="640" alt="09_dist_000000074177" src="https://github.com/user-attachments/assets/628540af-6407-4e93-b41e-9057671f49ca" />|<img width="64" height="64" alt="image" src="https://github.com/user-attachments/assets/23d932d4-ce48-421e-9d1f-0dd8386852a3" />|<img width="500" height="500" alt="08_dist_000000039322" src="https://github.com/user-attachments/assets/64c744c4-4425-46da-a53b-ffab2fd52060" />|

## Arch

<details><summary>Click to expand</summary>

<table>
 <tr>
 <th>ONNX</th>
 <th>LiteRT</th>
 </tr>
 <tr>
 <td valign="top"><img width="150" alt="ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost" src="https://github.com/user-attachments/assets/0f81e550-8e27-4f5a-b6a3-6ac79638dfa6" /></td>
 <td valign="top"><img width="150" alt="ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost_litert" src="https://github.com/user-attachments/assets/76809ab1-0ca9-4a2c-a796-db87c59ff076" /></td>
 </tr>
</table>

</details>

## Download all ONNX files at once
**I don't recommend it as it downloads a crazy amount of files.**

```bash
sudo apt update && sudo apt install -y gh
gh release download onnx -R PINTO0309/UHD
```

## Models

- Legacy models

<details><summary>Click to expand</summary>

- w/o ESE

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.38 M|0.18 G|0.40343|0.93 ms|5.6 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_quality_nopost.onnx)|
 |T|3.10 M|0.41 G|0.44529|1.50 ms|12.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_quality_nopost.onnx)|
 |S|5.43 M|0.71 G|0.44945|2.23 ms|21.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_quality_nopost.onnx)|
 |C|8.46 M|1.11 G|0.45005|2.66 ms|33.9 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_quality_nopost.onnx)|
 |M|12.15 M|1.60 G|0.44875|4.07 ms|48.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_quality_nopost.onnx)|
 |L|21.54 M|2.83 G|0.44686|6.23 ms|86.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_quality_nopost.onnx)|

- w ESE

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.45 M|0.18 G|0.41018|1.05 ms|5.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_se_64x64_quality_nopost.onnx)|
 |T|3.22 M|0.41 G|0.44130|1.27 ms|12.9 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_se_64x64_quality_nopost.onnx)|
 |S|5.69 M|0.71 G|0.46612|2.10 ms|22.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_se_64x64_quality_nopost.onnx)|
 |C|8.87 M|1.11 G|0.45095|2.86 ms|35.5 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_se_64x64_quality_nopost.onnx)|
 |M|12.74 M|1.60 G|0.46502|3.95 ms|51.0 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_se_64x64_quality_nopost.onnx)|
 |L|22.59 M|2.83 G|0.45787|6.52 ms|90.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_se_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_se_64x64_quality_nopost.onnx)|

- ESE + IoU-aware + Swish

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.60 M|0.20 G|0.42806|1.25 ms|6.5 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_se_iou_64x64_quality_nopost.onnx)|
 |T|3.56 M|0.45 G|0.46502|1.82 ms|14.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_se_iou_64x64_quality_nopost.onnx)|
 |S|6.30 M|0.79 G|0.47473|2.78 ms|25.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_se_iou_64x64_quality_nopost.onnx)|
 |C|9.81 M|1.23 G|0.46235|3.58 ms|39.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_se_iou_64x64_quality_nopost.onnx)|
 |M|14.09 M|1.77 G|0.46562|5.05 ms|56.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_se_iou_64x64_quality_nopost.onnx)|
 |L|24.98 M|3.13 G|0.47774|7.46 ms|100 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_se_iou_64x64_quality.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_se_iou_64x64_quality_nopost.onnx)|

- ESE + IoU-aware + ReLU

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.60 M|0.20 G|0.40910|0.63 ms|6.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_quality_relu_nopost.onnx)|
 |T|3.56 M|0.45 G|0.44618|1.08 ms|14.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_quality_relu_nopost.onnx)|
 |S|6.30 M|0.79 G|0.45776|1.71 ms|25.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_quality_relu_nopost.onnx)|
 |C|9.81 M|1.23 G|0.45385|2.51 ms|39.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_quality_relu_nopost.onnx)|
 |M|14.09 M|1.77 G|0.47468|3.54 ms|56.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_quality_relu_nopost.onnx)|
 |L|24.98 M|3.13 G|0.46965|6.14 ms|100 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_quality_relu.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_quality_relu_nopost.onnx)|

- ESE + IoU-aware + large-object-branch + ReLU

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.98 M|0.22 G|0.40903|0.77 ms|8.0 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_loese_nopost.onnx)|
 |T|4.40 M|0.49 G|0.46170|1.40 ms|17.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_loese_nopost.onnx)|
 |S|7.79 M|0.87 G|0.45860|2.30 ms|31.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_loese_nopost.onnx)|
 |C|12.13 M|1.35 G|0.47518|2.83 ms|48.6 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_loese_nopost.onnx)|
 |M|17.44 M|1.94 G|0.45816|4.37 ms|69.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_loese_nopost.onnx)|
 |L|30.92 M|3.44 G|0.48243|7.40 ms|123.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_loese.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_loese_nopost.onnx)|

- **[For long distances and extremely small objects]** ESE + IoU-aware + ReLU + Distillation

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.60 M|0.20 G|0.55224|0.63 ms|6.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_distill_nopost.onnx)|
 |T|3.56 M|0.45 G|0.56040|1.08 ms|14.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_distill_nopost.onnx)|
 |S|6.30 M|0.79 G|0.57361|1.71 ms|25.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_distill_nopost.onnx)|
 |C|9.81 M|1.23 G|0.56183|2.51 ms|39.3 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_distill_nopost.onnx)|
 |M|14.09 M|1.77 G|0.57666|3.54 ms|56.4 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_distill_nopost.onnx)|

- **[For short/medium distance]** ESE + IoU-aware + large-object-branch + ReLU + Distillation

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.98 M|0.22 G|0.54883|0.70 ms|8.0 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_loese_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_loese_distill_nopost.onnx)|
 |T|4.40 M|0.49 G|0.55663|1.18 ms|17.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_loese_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_loese_distill_nopost.onnx)|
 |S|7.79 M|0.87 G|0.57397|1.97 ms|31.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_loese_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_loese_distill_nopost.onnx)|
 |C|12.13 M|1.35 G|0.56768|2.74 ms|48.6 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_loese_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_loese_distill_nopost.onnx)|
 |M|17.44 M|1.94 G|0.57815|3.57 ms|69.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_loese_distill.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_loese_distill_nopost.onnx)|

- `torch_bilinear_dynamic` + No resizing required + Not suitable for quantization

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.98 M|0.22 G|0.55489|0.70 ms|8.0 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_torch_bilinear_dynamic_nopost.onnx)|
 |T|4.40 M|0.49 G|0.57824|1.18 ms|17.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_torch_bilinear_dynamic_nopost.onnx)|
 |S|7.79 M|0.87 G|0.58478|1.97 ms|31.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_torch_bilinear_dynamic_nopost.onnx)|
 |C|12.13 M|1.35 G|0.58459|2.74 ms|48.6 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_torch_bilinear_dynamic_nopost.onnx)|
 |M|17.44 M|1.94 G|0.59034|3.57 ms|69.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_torch_bilinear_dynamic_nopost.onnx)|
 |L|30.92 M|3.44 G|0.58929|7.16 ms|123.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_torch_bilinear_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_torch_bilinear_dynamic_nopost.onnx)|

- `torch_nearest_dynamic` + No resizing required + Suitable for quantization

|Variant|Params|FLOPs|mAP@0.5|Corei9 CPU inference latency|ONNX File size|ONNX|w/o post ONNX|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |N|1.98 M|0.22 G|0.53376|0.70 ms|8.0 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_torch_nearest_dynamic_nopost.onnx)|
 |T|4.40 M|0.49 G|0.55561|1.18 ms|17.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w96_64x64_torch_nearest_dynamic_nopost.onnx)|
 |S|7.79 M|0.87 G|0.56396|1.97 ms|31.2 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w128_64x64_torch_nearest_dynamic_nopost.onnx)|
 |C|12.13 M|1.35 G|0.56328|2.74 ms|48.6 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w160_64x64_torch_nearest_dynamic_nopost.onnx)|
 |M|17.44 M|1.94 G|0.57075|3.57 ms|69.8 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w192_64x64_torch_nearest_dynamic_nopost.onnx)|
 |L|30.92 M|3.44 G|0.56787|7.16 ms|123.7 MB|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_torch_nearest_dynamic.onnx)|[Download](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_torch_nearest_dynamic_nopost.onnx)|

</details>

- Variants
  ```
  R: ront, Y: yocto, Z: zepto, A: atto
  F: femto, P: pico, N: nano, T: tiny
  S: small, C: compact, M: medium, L: large
  ```
- `opencv_inter_nearest` + Optimized for OpenCV RGB downsampling + Suitable for quantization

- ONNX

|Var|Param|FLOPs|@0.5|Corei9 CPU latency|ONNX size|static|w/o post|dynamic|w/o post|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |R|0.13 M|0.01 G|0.21230|0.24 ms|863 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |Y|0.29 M|0.03 G|0.28664|0.28 ms|1.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |Z|0.51 M|0.05 G|0.32722|0.32 ms|2.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |A|0.78 M|0.08 G|0.43661|0.37 ms|3.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |F|1.12 M|0.12 G|0.47942|0.44 ms|4.5 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |P|1.52 M|0.17 G|0.51094|0.50 ms|6.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |N|1.98 M|0.22 G|0.55003|0.60 ms|8.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |T|2.49 M|0.28 G|0.56550|0.70 ms|10.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |S|3.07 M|0.34 G|0.57015|0.81 ms|12.3 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|
 |L|30.92 M|3.44 G|0.58399|7.16 ms|123.7 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_dynamic_nopost.onnx)|

- ESPDL INT8 (.espdl, .info, .json, anchors.npy, wh_scale.npy)

https://github.com/PINTO0309/esp-who/tree/custom/examples/ultra_lightweight_human_detection

|Var|ESPDL size|s3 Emphasis on speed|s3 latency|s3 Emphasis on precision|s3 latency|p4 static w/o post|p4 latency|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |R|218.3 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar.gz)|11.28 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar.gz)|20.05 ms|||
 |Y|380.2 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar.gz)|26.15 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar.gz)|28.51 ms|||
 |Z|601.1 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar.gz)|33.66 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar.gz)|48.57 ms|||
 |A|882.8 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar.gz)|127.51 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar.gz)|205.96 ms|||
 |F|1.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar_nocat.gz)|292.63 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar_nocat_espdl_highacc.gz)|-|||
 |P|1.6 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar_nocat.gz)|545.65 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar_nocat_espdl_highacc.gz)|-|||
 |N|2.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_static_nopost_nocat_espdl.tar_nocat.gz)|808.49 ms|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_static_nopost_nocat_espdl_highacc.tar_nocat_espdl_highacc.gz)|-|||

- ESP32-S3-EYE - Emphasis on precision `R`

https://github.com/user-attachments/assets/ad417641-0c06-4307-9224-a1bdb9402dbd

- `opencv_inter_nearest_yuv422` + Optimized for YUV422 + Suitable for quantization
  - `YUV422`
    ```python
    img_u8 = np.ones([64,64,3], dtype=np.uint8)
    yuyv = cv2.cvtColor(img_u8, cv2.COLOR_RGB2YUV_YUYV)
    print(yuyv.shape)
    (64, 64, 2)
    ```
  - With post-process model
    ```
    input_name.1: input_yuv422 shape: [1, 2, 64, 64] dtype: float32

output_name.1: score_classid_cxcywh shape: [1, 100, 6] dtype: float32
    ```
  - Without post-process model
    ```
    input_name.1: input_yuv422 shape: [1, 2, 64, 64] dtype: float32

output_name.1: txtywh_obj_quality_cls_x8 shape: [1, 56, 8, 8] dtype: float32
    output_name.2: anchors shape: [8, 2] dtype: float32
    output_name.3: wh_scale shape: [8, 2] dtype: float32

https://github.com/PINTO0309/UHD/blob/e0bbfe69afa0da4f83cf1f09b530a500bcd2d685/demo_uhd.py#L203-L301

score_mode = obj_quality_cls:
      score = sigmoid(obj) * sigmoid(quality) ** quality_power * sigmoid(cls)
    score_mode = quality_cls:
      score = sigmoid(quality) ** quality_power * sigmoid(cls)
    score_mode = obj_cls:
      score = sigmoid(obj) * sigmoid(cls)
    quality_power = --quality-power (default 1.0)
    note: if the quality head is absent, quality is ignored and all modes reduce to sigmoid(obj) * sigmoid(cls)
    cx = (sigmoid(tx)+gx)/w
    cy = (sigmoid(ty)+gy)/h
    bw = anchor_w*softplus(tw)*wh_scale
    bh = anchor_h*softplus(th)*wh_scale
    boxes = (cx±bw/2, cy±bh/2)
    ```
  - ONNX

|Var|Param|FLOPs|@0.5|Corei9 CPU latency|ONNX size|static|w/o post|dynamic|w/o post|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |R|0.13 M|0.01 G|0.22382|0.34 ms|863 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |Y|0.29 M|0.03 G|0.29606|0.38 ms|1.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |Z|0.51 M|0.05 G|0.36843|0.43 ms|2.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |A|0.78 M|0.08 G|0.42872|0.48 ms|3.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |F|1.12 M|0.12 G|0.49098|0.54 ms|4.5 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |P|1.52 M|0.17 G|0.52665|0.63 ms|6.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |N|1.98 M|0.22 G|0.54942|0.70 ms|8.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |T|2.49 M|0.28 G|0.56300|0.83 ms|10.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |S|3.07 M|0.34 G|0.57338|0.91 ms|12.3 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|
 |L|30.92 M|3.44 G|0.58642|7.16 ms|123.7 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_yuv422_distill_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_yuv422_distill_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_yuv422_distill_dynamic_nopost.onnx)|

- Input image 480x360 -> OpenCV INTER_NEAREST -> 64x64 -> YUV422 (packed: YUY2/YUYV)

|100%|800% zoom|
 |:-:|:-:|
 |<img width="64" height="64" alt="resized_64x64_nearest" src="https://github.com/user-attachments/assets/fd4d0ee3-1d6a-4819-a324-05f406de999c" />|<img width="515" height="515" alt="image" src="https://github.com/user-attachments/assets/ec62913a-236b-4c94-8ce7-e20ff67b69f8" />|

- `Y` detection sample

- `F` detection sample

- `N` detection sample

- `S` detection sample

- ESPDL INT8 (.espdl, .info, .json, anchors.npy, wh_scale.npy)

|Var|ESPDL size|static w/o post s3|static w/o post p4|
 |:-:|:-:|:-:|:-:|
 |R|218.3 KB|||||
 |Y|380.2 KB|||||
 |Z|601.1 KB|||||
 |A|882.8 KB|||||
 |F|1.2 MB|||||
 |P|1.6 MB|||||
 |N|2.1 MB|||||

- `opencv_inter_nearest_y` + Optimized for Y (Luminance) only + Suitable for quantization
  - `Y`
    ```python
    img_u8 = np.ones([64,64,3], dtype=np.uint8)
    yuv = cv2.cvtColor(img_u8, cv2.COLOR_RGB2YUV)
    y = yuv[..., 0:1]
    print(y.shape)
    (64, 64, 1)
    ```
  - With post-process model
    ```
    input_name.1: input_y shape: [1, 1, 64, 64] dtype: float32

output_name.1: score_classid_cxcywh shape: [1, 100, 6] dtype: float32
    ```
  - Without post-process model
    ```
    input_name.1: input_yuv422 shape: [1, 1, 64, 64] dtype: float32

output_name.1: txtywh_obj_quality_cls_x8 shape: [1, 56, 8, 8] dtype: float32
    output_name.2: anchors shape: [8, 2] dtype: float32
    output_name.3: wh_scale shape: [8, 2] dtype: float32

https://github.com/PINTO0309/UHD/blob/e0bbfe69afa0da4f83cf1f09b530a500bcd2d685/demo_uhd.py#L203-L301

|Var|Param|FLOPs|@0.5|CPU latency|ONNX size|static|w/o post|dynamic|w/o post|
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |R|0.13 M|0.01 G|0.21863|0.34 ms|863 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |Y|0.29 M|0.03 G|0.29523|0.38 ms|1.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |Z|0.51 M|0.05 G|0.33219|0.43 ms|2.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |A|0.78 M|0.08 G|0.41927|0.48 ms|3.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |F|1.12 M|0.12 G|0.46622|0.54 ms|4.5 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |P|1.52 M|0.17 G|0.51670|0.63 ms|6.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |N|1.98 M|0.22 G|0.54105|0.70 ms|8.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |T|2.49 M|0.28 G|0.55636|0.83 ms|10.0 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w72_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |S|3.07 M|0.34 G|0.56824|0.91 ms|12.3 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w80_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|
 |L|30.92 M|3.44 G|0.58164|7.16 ms|123.7 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_y_static.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_y_static_nopost.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_y_dynamic.onnx)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w256_64x64_opencv_inter_nearest_y_dynamic_nopost.onnx)|

- ESPDL INT8 (.espdl, .info, .json, anchors.npy, wh_scale.npy)

|Var|ESPDL size|s3 Emphasis on speed|s3 Emphasis on precision|
 |:-:|:-:|:-:|:-:|
 |R|218.3 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl_highacc.tar.gz)|
 |Y|380.2 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w24_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl_highacc.tar.gz)|
 |Z|601.1 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl_highacc.tar.gz)|
 |A|882.8 KB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w40_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl_highacc.tar.gz)|
 |F|1.2 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w48_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|-|
 |P|1.6 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w56_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|-|
 |N|2.1 MB|[DL](https://github.com/PINTO0309/UHD/releases/download/onnx/ultratinyod_res_anc8_w64_64x64_opencv_inter_nearest_y_static_nopost_nocat_espdl.tar.gz)|-|

- `opencv_inter_nearest_y_tri` + Optimized for Y (Luminance) only + Y ternarization + Suitable for quantization
 - `opencv_inter_nearest_y_tri` uses fixed Y thresholds (1/3, 2/3) to quantize to 3 levels: 0.0, 0.5, 1.0.
 - `Y_TRI`
 ```python
 img_u8 = np.ones([64,64,3], dtype=np.uint8)
 yuv = cv2.cvtColor(img_u8, cv2.COLOR_RGB2YUV)
 y = yuv[..., 0:1]
 y = y / 255.0
 y = np.clip(y, 0.0, 1.0)
 out = np.zeros_like(y, dtype=np.float32)
 mid_mask = (y >= t1) & (y < t2)
 out[mid_mask] = 0.5
 out[y >= t2] = 1.0
 print(out.shape)
 (64, 64, 1)
 ```
 - With post-process model
 ```
 input_name.1: input_y_tri shape: [1, 1, 64, 64] dtype: float32

output_name.1: score_classid_cxcywh shape: [1, 100, 6] dtype: float32
    ```
  - Without post-process model
    ```
    input_name.1: input_y_tri shape: [1, 1, 64, 64] dtype: float32

output_name.1: txtywh_obj_quality_cls_x8 shape: [1, 56, 8, 8] dtype: float32
    output_name.2: anchors shape: [8, 2] dtype: float32
    output_name.3: wh_scale shape: [8, 2] dtype: float32

https://github.com/PINTO0309/UHD/blob/e0bbfe69afa0da4f83cf1f09b530a500bcd2d685/demo_uhd.py#L203-L301

**WIP**

- `opencv_inter_nearest_y_bin` + Optimized for Y (Luminance) only + Y binarization + Suitable for quantization
  - `opencv_inter_nearest_y_bin` uses a fixed Y threshold (0.5) to binarize to 0.0/1.0.
  - `Y_BIN`
    ```python
    img_u8 = np.ones([64,64,3], dtype=np.uint8)
    yuv = cv2.cvtColor(img_u8, cv2.COLOR_RGB2YUV)
    y = yuv[..., 0:1]
    y = y / 255.0
    y = np.clip(y, 0.0, 1.0)
    out = (y >= threshold).astype(np.float32)
    print(out.shape)
    (64, 64, 1)
    ```
  - With post-process model
    ```
    input_name.1: input_y_bin shape: [1, 1, 64, 64] dtype: float32

output_name.1: score_classid_cxcywh shape: [1, 100, 6] dtype: float32
    ```
  - Without post-process model
    ```
    input_name.1: input_y_bin shape: [1, 1, 64, 64] dtype: float32

output_name.1: txtywh_obj_quality_cls_x8 shape: [1, 56, 8, 8] dtype: float32
    output_name.2: anchors shape: [8, 2] dtype: float32
    output_name.3: wh_scale shape: [8, 2] dtype: float32

https://github.com/PINTO0309/UHD/blob/e0bbfe69afa0da4f83cf1f09b530a500bcd2d685/demo_uhd.py#L203-L301

**WIP**

## Inference

> [!CAUTION]
> If you preprocess your images and resize them to 64x64 with OpenCV or similar, use `Nearest` mode.

<details><summary>Click to expand</summary>

```bash
usage: demo_uhd_lite.py
[-h]
(--images IMAGES | --camera CAMERA)
--onnx ONNX
[--output OUTPUT]
[--img-size IMG_SIZE]
[--conf-thresh CONF_THRESH]
[--record RECORD]
[--use-nms]
[--nms-iou NMS_IOU]
[--topk TOPK]
[--topk-before-conf]

UltraTinyOD ONNX/LiteRT demo (CPU).

options:
  -h, --help            show this help message and exit
  --images IMAGES       Directory with images to run batch inference.
  --camera CAMERA       USB camera id for realtime inference.
  --onnx ONNX           Path to ONNX model (CPU).
  --output OUTPUT       Output directory for image mode.
  --img-size IMG_SIZE   Input size HxW, e.g., 64x64.
  --conf-thresh CONF_THRESH
                        Confidence threshold.
  --record RECORD       MP4 path for recording in camera mode.
  --use-nms             Apply IoU NMS (single class).
  --nms-iou NMS_IOU     IoU threshold for NMS.
  --topk TOPK           Keep top-K boxes before NMS.
  --topk-before-conf    Apply top-K before confidence threshold.

```

```bash
python demo_uhd_lite.py \
--onnx ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx \
--camera 0 \
--use-nms
```

</details>

## Training Examples (full CLI)

UltraTinyOD (anchor-only, stride 8; `--cnn-width` controls stem width):

<details><summary>64x64 - single-class - small</summary>

```bash
# Use large-obj-branch
# High accuracy but slightly slow
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
LR=0.03
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/train.txt \
--val-list data/wholebody34/val.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_s.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg17 \
--batch-size 64 \
--num-workers 12 \
--epochs 300 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-ema \
--ema-decay 0.9999 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--utod-large-obj-branch \
--utod-large-obj-depth 1 \
--utod-large-obj-ch-scale 1.25 \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls
```
```bash
# Disable large-obj-branch
# Less accurate but slightly faster
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
LR=0.03
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/train.txt \
--val-list data/wholebody34/val.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_s.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg18_nolo \
--batch-size 64 \
--num-workers 12 \
--epochs 300 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-ema \
--ema-decay 0.9999 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls
```

</details>

<details><summary>64x64 - single-class - Large</summary>

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=512
LR=0.0007
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/train.txt \
--val-list data/wholebody34/val.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_l.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg16 \
--batch-size 64 \
--num-workers 12 \
--epochs 300 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--utod-large-obj-branch \
--utod-large-obj-depth 1 \
--utod-large-obj-ch-scale 1.25 \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls
```

</details>

<details><summary>64x64 - single-class - Large - self-distillation - fine-tuning</summary>

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=512
LR=0.0007
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/trainft.txt \
--val-list data/wholebody34/valft.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_l.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg16_distill \
--ckpt runs/ultratinyod_anc8_w512_dw_64x64_lr0.0007_opencv_inter_nearest_cg16/best_utod_0287_map_0.52045.pt \
--batch-size 64 \
--num-workers 12 \
--epochs 100 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--utod-large-obj-branch \
--utod-large-obj-depth 1 \
--utod-large-obj-ch-scale 1.25 \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls \
--teacher-ckpt runs/ultratinyod_anc8_w512_dw_64x64_lr0.0007_opencv_inter_nearest_cg16/best_utod_0287_map_0.52045.pt \
--teacher-arch ultratinyod \
--distill-temperature 1.5 \
--distill-cosine \
--distill-box-l1 0.05 \
--distill-obj 0.1 \
--distill-quality 0.1
```

</details>

<details><summary>64x64 - single-class - small - distillation</summary>

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
LR=0.00001
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/trainft.txt \
--val-list data/wholebody34/valft.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_s.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg17_distill \
--ckpt runs/ultratinyod_anc8_w32_dw_64x64_lr0.03_opencv_inter_nearest_cg17/best_utod_0211_map_0.27713.pt \
--batch-size 64 \
--num-workers 12 \
--epochs 100 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--utod-large-obj-branch \
--utod-large-obj-depth 1 \
--utod-large-obj-ch-scale 1.25 \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls \
--teacher-ckpt runs/ultratinyod_anc8_w512_dw_64x64_lr0.0007_opencv_inter_nearest_cg16_distill/best_utod_0100_map_0.54755.pt \
--teacher-arch ultratinyod \
--distill-temperature 1.0 \
--distill-cosine \
--distill-box-l1 0.05 \
--distill-obj 0.1 \
--distill-quality 0.1
```
```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
LR=0.00001
RESIZEMODE=opencv_inter_nearest
uv run python train.py \
--arch ultratinyod \
--train-list data/wholebody34/trainft.txt \
--val-list data/wholebody34/valft.txt \
--aug-config uhd/aug.yaml \
--val-aug-config uhd/aug_val_s.yaml \
--img-size ${SIZE} \
--exp-name ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_dw_${SIZE}_lr${LR}_${RESIZEMODE}_cg18_nolo_distill \
--ckpt runs/ultratinyod_anc8_w32_dw_64x64_lr0.03_opencv_inter_nearest_cg18_nolo/best_utod_0206_map_0.27951.pt \
--batch-size 64 \
--num-workers 12 \
--epochs 100 \
--lr ${LR} \
--weight-decay 0.0001 \
--device cuda \
--classes 0 \
--cnn-width ${CNNWIDTH} \
--anchors "0.043750,0.094437 0.141667,0.394426 0.193750,0.540625 0.254172,0.636096 0.320819,0.708333 0.414578,0.762981 0.531250,0.833326 0.739375,0.916667" \
--num-anchors ${ANCHOR} \
--iou-loss ciou \
--conf-thresh 0.15 \
--use-batchnorm \
--utod-conv dw \
--utod-residual \
--utod-sppf-scale conv \
--use-improved-head \
--use-iou-aware-head \
--activation relu \
--${RESIZEMODE} \
--loss-weight-box 1.0 \
--loss-weight-obj 16.0 \
--loss-weight-quality 16.0 \
--obj-loss bce \
--obj-target iou \
--disable-cls \
--teacher-ckpt runs/ultratinyod_anc8_w512_dw_64x64_lr0.0007_opencv_inter_nearest_cg16_distill/best_utod_0100_map_0.54755.pt \
--teacher-arch ultratinyod \
--distill-temperature 1.0 \
--distill-cosine \
--distill-box-l1 0.05 \
--distill-obj 0.1 \
--distill-quality 0.1
```

</details>

## Validation-only Example

<details><summary>Click to expand</summary>

Example of running only validation on a trained checkpoint:

```bash
uv run python train.py \
--arch ultratinyod \
--img-size 64x64 \
--cnn-width 256 \
--classes 0 \
--conf-thresh 0.15 \
--ckpt runs/ultratinyod_res_anc8_w256_64x64_lr0.0003/best_utod_0297_map_0.44299.pt \
--val-only \
--use-ema \
--with-heatmap-mean
```

When `--with-heatmap-mean` or `--with-heatmap-sum` is set, validation sample outputs are written under
`runs/<exp-name>/val_only/<sample_name>/`, containing the usual detection PNG plus per-layer heatmap PNGs
overlayed on the detection image.

</details>

## CLI parameters

<details><summary>Click to expand</summary>

| Parameter | Description | Default |
| --- | --- | --- |
| `--arch` | Model architecture: `cnn`, `transformer`, or anchor-only `ultratinyod`. | `cnn` |
| `--image-dir` | Directory containing images and YOLO txt labels (required when not using `--train-list`/`--val-list`). | `None` |
| `--train-list` | Optional train list file (one image path per line); requires `--val-list` and ignores split ratios. | `None` |
| `--val-list` | Optional validation list file (one image path per line); requires `--train-list` and ignores split ratios. | `None` |
| `--train-split` | Fraction of data used for training. | `0.8` |
| `--val-split` | Fraction of data used for validation. | `0.2` |
| `--img-size` | Input size `HxW` (e.g., `64x64`). | `64x64` |
| `--resize-mode` | Resize mode for training preprocessing: `torch_bilinear`, `torch_nearest`, `opencv_inter_linear`, `opencv_inter_nearest`, `opencv_inter_nearest_y`, `opencv_inter_nearest_y_bin`, `opencv_inter_nearest_y_tri`, `opencv_inter_nearest_yuv422`. | `torch_bilinear` |
| `--torch_bilinear` | Shortcut for `--resize-mode torch_bilinear`. | `False` |
| `--torch_nearest` | Shortcut for `--resize-mode torch_nearest`. | `False` |
| `--opencv_inter_linear` | Shortcut for `--resize-mode opencv_inter_linear`. | `False` |
| `--opencv_inter_nearest` | Shortcut for `--resize-mode opencv_inter_nearest`. | `False` |
| `--opencv_inter_nearest_y` | Shortcut for `--resize-mode opencv_inter_nearest_y`. | `False` |
| `--opencv_inter_nearest_y_bin` | Shortcut for `--resize-mode opencv_inter_nearest_y_bin`. | `False` |
| `--opencv_inter_nearest_y_tri` | Shortcut for `--resize-mode opencv_inter_nearest_y_tri`. | `False` |
| `--opencv_inter_nearest_yuv422` | Shortcut for `--resize-mode opencv_inter_nearest_yuv422`. | `False` |
| `--exp-name` | Experiment name; logs saved under `runs/<exp-name>`. | `default` |
| `--batch-size` | Batch size. | `64` |
| `--epochs` | Number of epochs. | `100` |
| `--resume` | Checkpoint to resume training (loads optimizer/scheduler). | `None` |
| `--ckpt` | Initialize weights from checkpoint (no optimizer state). | `None` |
| `--ckpt-non-strict` | Load `--ckpt` with `strict=False` (ignore missing/unexpected keys). | `False` |
| `--val-only` | Run validation only with `--ckpt` or `--resume` weights and exit. | `False` |
| `--val-count` | Limit number of validation images when using `--val-only`. | `None` |
| `--with-heatmap-mean` | Save per-layer heatmaps using channel mean and overlay them on detection images. | `False` |
| `--with-heatmap-sum` | Save per-layer heatmaps using channel sum and overlay them on detection images. | `False` |
| `--use-improved-head` | UltraTinyOD only: enable quality-aware head (IoU-aware obj, IoU score branch, learnable WH scale, extra context). | `False` |
| `--use-iou-aware-head` | UltraTinyOD head: task-aligned IoU-aware scoring (quality*cls) with split towers. | `False` |
| `--quality-power` | Exponent for quality score when using IoU-aware head scoring. | `1.0` |
| `--score-mode` | Score composition mode for anchor head (`obj_quality_cls`, `quality_cls`, `obj_cls`, `obj_quality`, `quality`, `obj`). Normally overridden by checkpoint meta; with `--val-only` and explicit `--score-mode`, CLI takes priority. | `None` |
| `--teacher-ckpt` | Teacher checkpoint path for distillation. | `None` |
| `--teacher-arch` | Teacher architecture override. | `None` |
| `--teacher-num-queries` | Teacher DETR queries. | `None` |
| `--teacher-d-model` | Teacher model dimension. | `None` |
| `--teacher-heads` | Teacher attention heads. | `None` |
| `--teacher-layers` | Teacher encoder/decoder layers. | `None` |
| `--teacher-dim-feedforward` | Teacher FFN dimension. | `None` |
| `--teacher-use-skip` | Force teacher skip connections on. | `False` |
| `--teacher-activation` | Teacher activation (`relu`/`swish`). | `None` |
| `--teacher-use-fpn` | Force teacher FPN on. | `False` |
| `--teacher-backbone` | Teacher backbone checkpoint for feature distillation. | `None` |
| `--teacher-backbone-arch` | Teacher backbone architecture hint. | `None` |
| `--teacher-backbone-norm` | Teacher backbone input normalization. | `imagenet` |
| `--distill-kl` | KL distillation weight (transformer). | `0.0` |
| `--distill-box-l1` | Box L1 distillation weight (transformer). | `0.0` |
| `--distill-obj` | Objectness distillation weight (anchor head). | `0.0` |
| `--distill-quality` | Quality-score distillation weight (anchor head). | `0.0` |
| `--distill-cosine` | Cosine ramp-up of distillation weights. | `False` |
| `--distill-temperature` | Teacher logits temperature. | `1.0` |
| `--distill-feat` | Feature-map distillation weight (CNN only). | `0.0` |
| `--lr` | Learning rate. | `0.001` |
| `--weight-decay` | Weight decay. | `0.0001` |
| `--optimizer` | Optimizer (`adamw` or `sgd`). | `adamw` |
| `--grad-clip-norm` | Global gradient norm clip; set `0` to disable. | `5.0` |
| `--num-workers` | DataLoader workers. | `8` |
| `--device` | Device: `cuda` or `cpu`. | `cuda` if available |
| `--seed` | Random seed. | `42` |
| `--log-interval` | Steps between logging to progress bar. | `10` |
| `--eval-interval` | Epoch interval for evaluation. | `1` |
| `--conf-thresh` | Confidence threshold for decoding. | `0.3` |
| `--topk` | Top-K for CNN decoding. | `50` |
| `--use-amp` | Enable automatic mixed precision. | `False` |
| `--qat` | Enable QAT via `torch.ao.quantization`. | `False` |
| `--qat-backend` | Quantization backend for QAT (`fbgemm` or `qnnpack`). | `fbgemm` |
| `--qat-fuse` | Fuse Conv+BN(+ReLU) before QAT. | `False` |
| `--qat-disable-observer-epoch` | Epoch (1-based) to disable observers (`0` to skip). | `2` |
| `--qat-freeze-bn-epoch` | Epoch (1-based) to freeze BN stats (`0` to skip). | `3` |
| `--w-bits` | Fake-quant bits for weights (UltraTinyOD only, `0` disables). | `0` |
| `--a-bits` | Fake-quant bits for activations (UltraTinyOD only, `0` disables). | `0` |
| `--quant-target` | Quantization target for UltraTinyOD (`backbone`, `head`, `both`, `none`). | `both` |
| `--lowbit-quant-target` | Low-bit quantization target (defaults to `--quant-target`). | `None` |
| `--lowbit-w-bits` | Low-bit weight bits (defaults to `--w-bits`). | `None` |
| `--lowbit-a-bits` | Low-bit activation bits (defaults to `--a-bits`). | `None` |
| `--highbit-quant-target` | High-bit quantization target (`backbone`, `head`, `both`, `none`). | `none` |
| `--highbit-w-bits` | High-bit weight bits. | `8` |
| `--highbit-a-bits` | High-bit activation bits. | `8` |
| `--aug-config` | YAML for augmentations (applied in listed order). | `uhd/aug.yaml` |
| `--use-ema` | Enable EMA of model weights for evaluation/checkpointing. | `False` |
| `--ema-decay` | EMA decay factor (ignored if EMA disabled). | `0.9998` |
| `--coco-eval` | Run COCO-style evaluation. | `False` |
| `--coco-per-class` | Log per-class COCO AP when COCO eval is enabled. | `False` |
| `--classes` | Comma-separated target class IDs. | `0` |
| `--activation` | Activation function (`relu` or `swish`). | `swish` |
| `--cnn-width` | Width multiplier for CNN backbone. | `32` |
| `--backbone` | Optional lightweight CNN backbone (`microcspnet`, `ultratinyresnet`, `enhanced-shufflenet`, or `none`). | `None` |
| `--backbone-channels` | Comma-separated channels for `ultratinyresnet` (e.g., `16,32,48,64`). | `None` |
| `--backbone-blocks` | Comma-separated residual block counts per stage for `ultratinyresnet` (e.g., `1,2,2,1`). | `None` |
| `--backbone-se` | Apply SE/eSE on backbone output (custom backbones only). | `none` |
| `--backbone-skip` | Add long skip fusion across custom backbone stages (ultratinyresnet). | `False` |
| `--backbone-skip-cat` | Use concat+1x1 fusion for long skips (ultratinyresnet); implies `--backbone-skip`. | `False` |
| `--backbone-skip-shuffle-cat` | Use stride+shuffle concat fusion for long skips (ultratinyresnet); implies `--backbone-skip`. | `False` |
| `--backbone-skip-s2d-cat` | Use space-to-depth concat fusion for long skips (ultratinyresnet); implies `--backbone-skip`. | `False` |
| `--backbone-fpn` | Enable a tiny FPN fusion inside custom backbones (ultratinyresnet). | `False` |
| `--backbone-out-stride` | Override custom backbone output stride (e.g., `8` or `16`). | `None` |
| `--use-skip` | Enable skip-style fusion in the CNN head (sums pooled shallow features into the final stage). Stored in checkpoints and restored on resume. | `False` |
| `--utod-residual` | Enable residual skips inside the UltraTinyOD backbone. | `False` |
| `--utod-head-ese` | UltraTinyOD head: apply lightweight eSE on shared features. | `False` |
| `--utod-conv` | UltraTinyOD conv mode (`dw` depthwise separable, `std` standard conv). | `dw` |
| `--utod-sppf-scale` | UltraTinyOD SPPF-min: per-branch scale matching before concat (`none`, `bn`, `conv`). | `none` |
| `--utod-context-rfb` | UltraTinyOD head: add a receptive-field block (dilated + wide depthwise) before prediction layers. | `False` |
| `--utod-context-dilation` | Dilation used in UltraTinyOD receptive-field block (only when `--utod-context-rfb`). | `2` |
| `--utod-large-obj-branch` | UltraTinyOD head: add a downsampled large-object refinement branch (no FPN). | `False` |
| `--utod-large-obj-depth` | Number of depthwise blocks in the large-object branch (only when `--utod-large-obj-branch`). | `2` |
| `--utod-large-obj-ch-scale` | Channel scale for the large-object branch (relative to head channels). | `1.0` |
| `--utod-quant-arch` | UltraTinyOD quantization-robust architecture mode: `0` off, `1` box stage2 residual gain, `2` box stage2 low-rank pw, `3` split `box_out` (`xy/wh`), `4` box activation clip (ReLU6), `5` gated large-object fusion, `6` (`1+5`), `7` (`1+3`), `8` (`2+3`), `9` box stage1+stage2 residual gain, `10` backbone stage1+stage2(+stage3/4) residualized, `11` (`1+10`). (`1`/`2`/`6`/`7`/`8`/`9`/`11` are designed for `--use-iou-aware-head`, `5`/`6` for `--utod-large-obj-branch`) | `0` |
| `--use-anchor` | Use anchor-based head for CNN (YOLO-style). | `False` |
| `--output-stride` | Final CNN feature stride (downsample factor). Supported: `4`, `8`, `16`. | `16` |
| `--anchors` | Anchor sizes as normalized `w,h` pairs (space separated). | `""` |
| `--auto-anchors` | Compute anchors from training labels when using anchor head. | `False` |
| `--auto-anchors-alg` | Auto-anchors algorithm (`kmeans`, `logkmeans`, `stratified`, `stratified_large`, `sml_fixed`). | `kmeans` |
| `--auto-anchors-plot` | Save a width/height distribution plot used for auto-anchors. | `False` |
| `--auto-anchors-plot-path` | Output path for auto-anchors plot (default `runs/EXP/auto_anchors_wh_<alg>.png`). | `None` |
| `--num-anchors` | Number of anchors to use when auto-computing. | `3` |
| `--iou-loss` | IoU loss type for anchor head (`iou`, `giou`, or `ciou`). | `giou` |
| `--anchor-assigner` | Anchor assigner strategy (`legacy`, `simota`). | `legacy` |
| `--anchor-cls-loss` | Anchor classification loss (`bce`, `vfl`, `ce`). | `bce` |
| `--disable-cls` | Disable cls branch entirely (classless anchor head; intended for single-class datasets). | `False` |
| `--loss-weight-box` | Loss weight for anchor box regression. | `1.0` |
| `--loss-weight-obj` | Loss weight for anchor objectness. | `1.0` |
| `--loss-weight-cls` | Loss weight for anchor classification. | `1.0` |
| `--loss-weight-quality` | Loss weight for anchor quality head. | `1.0` |
| `--obj-loss` | Objectness loss type for anchor head (`bce`, `smoothl1`). | `bce` |
| `--obj-target` | Objectness target for anchor head (`auto`, `binary`, `iou`). | `auto` |
| `--multi-label-mode` | Multi-label mode for anchor heads (`none`, `single`, `separate`). | `none` |
| `--multi-label-det-classes` | Comma-separated class ids for detection head when `--multi-label-mode separate`. | `None` |
| `--multi-label-attr-classes` | Comma-separated class ids for attribute head when `--multi-label-mode separate`. | `None` |
| `--multi-label-attr-weight` | Loss weight for attribute head when `--multi-label-mode separate`. | `1.0` |
| `--simota-topk` | Top-K IoUs for dynamic-k in SimOTA. | `10` |
| `--last-se` | Apply SE/eSE only on the last CNN block. | `none` |
| `--use-batchnorm` | Enable BatchNorm layers during training/export. | `False` |
| `--last-width-scale` | Channel scale for last CNN block (e.g., `1.25`). | `1.0` |
| `--num-queries` | Transformer query count. | `10` |
| `--d-model` | Transformer model dimension. | `64` |
| `--heads` | Transformer attention heads. | `4` |
| `--layers` | Transformer encoder/decoder layers. | `3` |
| `--dim-feedforward` | Transformer feedforward dimension. | `128` |
| `--use-fpn` | Enable simple FPN for transformer backbone. | `False` |

- When `--train-list`/`--val-list` are not provided, the split from `--train-split`/`--val-split` is written to `train.txt` and `val.txt` in the parent directory of `--image-dir`.
- Tiny CNN backbones (`--backbone`, optional; default keeps the original built-in CNN):
  - `microcspnet`: CSP-tiny style stem (16/32/64/128) compressed to 64ch, stride 8 output.
  - `ultratinyresnet`: 16→24→32→48 channel ResNet-like stack with three downsample steps (stride 8). Channel widths and blocks per stage can be overridden via `--backbone-channels` / `--backbone-blocks`; optional long skips across stages via `--backbone-skip`; optional lightweight FPN fusion via `--backbone-fpn`.
  - `enhanced-shufflenet`: Enhanced ShuffleNetV2+ inspired (arXiv:2111.00902) with progressive widening and doubled refinements, ending at ~128ch, stride 8.
- All custom backbones can optionally apply SE/eSE on the backbone output via `--backbone-se {none,se,ese}`.

</details>

## Augmentation via YAML

<details><summary>Click to expand</summary>

- Specify a YAML file with `--aug-config` to run the `data_augment:` entries in the listed order (e.g., `--aug-config uhd/aug.yaml`).
- Supported ops (examples): Mosaic / MixUp / CopyPaste / HorizontalFlip (class_swap_map supported) / VerticalFlip / RandomScale / Translation / RandomCrop / RandomResizedCrop / RandomBrightness / RandomContrast / RandomSaturation / RandomHSV / RandomPhotometricDistort / Blur / MedianBlur / MotionBlur / GaussianBlur / GaussNoise / ImageCompression / ISONoise / RandomRain / RandomFog / RandomSunFlare / CLAHE / ToGray / RemoveOutliers.
- If `prob` is provided, it is used as the apply probability; otherwise defaults are used (most are 0, RandomPhotometricDistort defaults to 0.5). Unknown keys are ignored.

</details>

## Loss terms (CNN / CenterNet)

<details><summary>Click to expand</summary>

- `loss`: total loss (`hm + off + wh`)
- `hm`: focal loss on center heatmap
- `off`: L1 loss on center offsets (within-cell quantization correction)
- `wh`: L1 loss on width/height (feature-map scale)

</details>

## Loss terms (CNN / Anchor head, `--use-anchor`)

<details><summary>Click to expand</summary>

- `loss`: total anchor loss (`box + obj + cls` [+ `quality`] when `--use-improved-head`)
- `obj`: BCE on objectness for each anchor location (positive vs. background)
- `cls`: Classification loss on per-class logits for positive anchors (BCE/VFL/CE per `--anchor-cls-loss`)
- `box`: (1 - IoU/GIoU/CIoU) on decoded boxes for positive anchors; IoU flavor set by `--iou-loss`
- `quality` (improved head only): BCE on IoU-linked quality logit; obj targetもIoUでスケールされる
- `--disable-cls` を指定すると cls ブランチ/損失は無効化され、予測は classless (`cls=0`) として扱われます。

</details>

## Loss terms (Transformer)

<details><summary>Click to expand</summary>

- `loss`: total loss (`cls + l1 + iou`)
- `cls`: cross-entropy for class vs. background
- `l1`: L1 loss on box coordinates
- `iou`: 1 - IoU for matched predictions

</details>

## The impact of image downsampling methods

<details><summary>Click to expand</summary>

PyTorch's Resize method is implemented using a downsampling method similar to PIL on the backend, but it is significantly different from OpenCV's downsampling implementation. Therefore, when downsampling images in preprocessing during training, it is important to note that the numerical characteristics of the images used by the model for training will be completely different depending on whether you use PyTorch's Resize method or OpenCV's Resize method. Below is the pixel-level error calculation when downsampling an image to 64x64 pixels. If the diff value is greater than 1.0, the images are completely different.

Therefore, it is easy to imagine that if the downsampling method used for preprocessing during learning is different from the downsampling method used during inference, the output inference results will be disastrous.

The internal workings of PyTorch's downsampling and PIL's downsampling are very similar but slightly different. When deploying and inferencing in Python and other environments, accuracy will be significantly degraded unless the model is deployed according to the following criteria. If you train using OpenCV's `cv2.INTER_LINEAR`, the model will never produce the correct output after preprocessing in PyTorch, TensorFlow, or ONNX other than OpenCV.

|Training|Deploy|
|:-|:-|
|When training while downsampling using PyTorch's Resize (`InterpolationMode.BILINEAR`)|Merge `Resize Linear + half-pixel` at the input of the ONNX model. This will result in the highest model accuracy. However, it will be limited to deployment on hardware, NPUs, and frameworks that support the resize operation of bilinear interpolation. It is not suitable for quantization.|
|When training while downsampling using PyTorch's Resize (`InterpolationMode.NEAREST`)|Merge `Resize Nearest` at the input of the ONNX model. It is the most versatile in terms of HW, NPU, and quantization deployment, but the accuracy of the model will be lower.|
|When training while downsampling using OpenCV's Resize (`cv2.INTER_NEAREST`)|Merge `Resize Nearest` at the input of the ONNX model or OpenCV `INTER_NEAREST`. Although the accuracy is low, it is highly versatile because the downsampling of images can be freely written on the program side. However, downsampling must be implemented manually.|

1. Error after PIL conversion when downsampling with PyTorch's Resize InterpolationMode.BILINEAR
 ```
 PyTorch(InterpolationMode.BILINEAR) -> Convert to PIL vs PyTorch Tensor(InterpolationMode.BILINEAR)
 max diff : 1
 mean diff : 0.4949
 std diff : 0.5000
 ```
2. Error when downsampling with PyTorch's Resize InterpolationMode.BILINEAR compared to downsampling with OpenCV's INTER_LINER
 ```
 PyTorch(InterpolationMode.BILINEAR) -> Convert to PIL vs OpenCV INTER_LINEAR
 max diff : 104
 mean diff : 10.2930
 std diff : 13.2792
 ```
3. Error when downsampling with PyTorch's Resize InterpolationMode.BILINEAR compared to downsampling with OpenCV's INTER_LINER
 ```
 PyTorch Tensor(InterpolationMode.BILINEAR) vs OpenCV INTER_LINEAR
 max diff : 104
 mean diff : 10.3336
 std diff : 13.2463
 ```
4. Accuracy and speed of each interpolation method when downsampling in OpenCV
 - Accuracy: `INTER_NEAREST < INTER_LINEAR < INTER_AREA`
 - Speed: `INTER_NEAREST > INTER_LINEAR > INTER_AREA`
 ```
 === Resize benchmark ===
 INTER_NEAREST : 0.0061 ms
 INTER_LINEAR : 0.0143 ms
 INTER_AREA : 0.3621 ms
 AREA / LINEAR ratio : 25.40x
 ```

</details>

## ONNX export

Notes:
- BatchNormalization is decomposed into Mul/Add after export by default. Disable with `--no-decompose-bn`.

### 1. No-Decode Post-Process - 1 class

<details><summary>Click to expand</summary>

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
RESIZEMODE=opencv_inter_nearest
CKPT=runs/ultratinyod_anc8_w32_dw_64x64_lr0.00001_opencv_inter_nearest_cg17_distill/best_utod_0100_map_0.31240.pt
uv run python export_onnx.py \
--checkpoint ${CKPT} \
--output ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
--opset 17 \
--no-merge-postprocess \
--noconcat_box_obj_quality_cls

uv run python uhd/quantize_onnx_model_for_esp32.py \
--dataset-type image \
--image-dir data/wholebody34/obj_train_data \
--resize-mode opencv_inter_nearest \
--onnx-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx \
--espdl-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.espdl \
--target esp32s3 \
--calib-algorithm kl \
--use-layerwise-equalization \
--int16-op-pattern /depthwiseconv/box_tower/box_tower.0/dw/conv/Conv \
--int16-op-pattern /box_tower/box_tower.0/dw/act/Relu \
--int16-op-pattern /box_tower/box_tower.0/pw/conv/Conv \
--int16-op-pattern /box_tower/box_tower.0/pw/act/Relu \
--int16-op-pattern /depthwiseconv/box_tower/box_tower.1/dw/conv/Conv \
--int16-op-pattern /box_tower/box_tower.1/dw/act/Relu \
--int16-op-pattern /box_tower/box_tower.1/pw/conv/Conv \
--int16-op-pattern /box_tower/box_tower.1/pw/act/Relu \
--int16-op-pattern /box_out/Conv \
--int16-op-pattern /quality_out/Conv
```
<img width="640" alt="image" src="https://github.com/user-attachments/assets/c6d5475a-c56a-4019-a24d-45543fafa4ff" />

</details>

### 2. Decoded Post-Process - 1 class

<details><summary>Click to expand</summary>

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=32
RESIZEMODE=opencv_inter_nearest
CKPT=runs/ultratinyod_anc8_w32_lo_64x64_lr0.03_opencv_inter_nearest_cg18/best_utod_0287_map_0.29338.pt
uv run python export_onnx.py \
--checkpoint ${CKPT} \
--output ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
--opset 17 \
--merge-primitive-postprocess

SIZE=64x64
ANCHOR=8
CNNWIDTH=40
RESIZEMODE=opencv_inter_nearest
CKPT=runs/ultratinyod_anc8_w40_lo_64x64_lr0.03_opencv_inter_nearest_cg18/best_utod_0232_map_0.31172.pt
uv run python export_onnx.py \
--checkpoint ${CKPT} \
--output ultratinyod_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
--opset 17 \
--merge-primitive-postprocess

</details>

### 3. No-Decode Post-Process - multi-class

<details><summary>Click to expand</summary>

- Export a checkpoint to ONNX (auto-detects arch from checkpoint unless overridden):
  ```bash
  SIZE=64x64
  ANCHOR=8
  CNNWIDTH=64
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w64_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls

SIZE=64x64
  ANCHOR=8
  CNNWIDTH=96
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w96_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls

SIZE=64x64
  ANCHOR=8
  CNNWIDTH=128
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w128_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls

SIZE=64x64
  ANCHOR=8
  CNNWIDTH=160
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w160_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls

SIZE=64x64
  ANCHOR=8
  CNNWIDTH=192
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w192_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls

SIZE=64x64
  ANCHOR=8
  CNNWIDTH=256
  RESIZEMODE=opencv_inter_nearest
  CKPT=runs/ultratinyod_res_anc8_w256_loese_64x64_lr0.005_impaug/best_utod_0001_map_0.00000.pt
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static.onnx \
  --opset 17
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic.onnx \
  --opset 17 \
  --dynamic-resize
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --noconcat_box_obj_quality_cls
  uv run python export_onnx.py \
  --checkpoint ${CKPT} \
  --output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_dynamic_nopost.onnx \
  --opset 17 \
  --no-merge-postprocess \
  --dynamic-resize \
  --noconcat_box_obj_quality_cls
  ```

</details>

## ONNX simple benchmark

<details><summary>Click to expand</summary>

```bash
uv run sit4onnx \
-if ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_dynamic_nopost.onnx \
-fs 1 3 480 640

INFO: file: ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_dynamic_nopost.onnx
INFO: providers: ['CPUExecutionProvider']
INFO: input_name.1: input_rgb shape: [1, 3, 480, 640] dtype: float32
INFO: test_loop_count: 10
INFO: total elapsed time:  3.634929656982422 ms
INFO: avg elapsed time per pred:  0.3634929656982422 ms
INFO: output_name.1: txtywh_obj_quality_cls_x8 shape: [1, 56, 8, 8] dtype: float32
INFO: output_name.2: anchors shape: [8, 2] dtype: float32
INFO: output_name.3: wh_scale shape: [8, 2] dtype: float32
```

</details>

## LiteRT (TFLite) quantization

Note that DepthwiseConv significantly reduces accuracy if per-channel quantization is not selected. Layerwise quantization is not suitable for this model.

<details><summary>Click to expand</summary>

```bash
# per-channel quantization
uv add tensorflow==2.19.0
uv run onnx2tf \
-i ultratinyod_res_anc8_w32_64x64_opencv_inter_nearest_dynamic_nopost.onnx \
-cotof \
-oiqt \
-qnm "[[[[0.0, 0.0, 0.0]]]]" \
-qns "[[[[1.0, 1.0, 1.0]]]]"
```

</details>

## ESP-DL Quantization

This section separates QAT and PTQ workflows for ESP-DL.

### 1. QAT (Quantization-Aware Training)

`uhd/uhd_qat.py` performs Quantization-Aware Training on a **raw** UltraTinyOD ONNX graph
and exports `.espdl`/`.native` per epoch.

#### 1-1. Export a raw ONNX (no postprocess)

<details><summary>Click to expand</summary>

QAT expects raw head outputs. Export with `--no-merge-postprocess` and `--noconcat_box_obj_quality_cls`. Quantization is possible even if you exported ONNX without specifying `--noconcat_box_obj_quality_cls`, but if you want to minimize the loss of accuracy due to quantization, it is better to export to ONNX with this option specified.

```bash
SIZE=64x64
ANCHOR=8
CNNWIDTH=16
RESIZEMODE=opencv_inter_nearest
CKPT=runs/17/ultratinyod_res_anc8_w16_loese_64x64_lr0.000001_opencv_inter_nearest_distill/best_utod_0008_map_0.21863.pt

uv run python export_onnx.py \
--checkpoint ${CKPT} \
--output ultratinyod_res_anc${ANCHOR}_w${CNNWIDTH}_${SIZE}_${RESIZEMODE}_static_nopost_nocat.onnx \
--opset 17 \
--no-merge-postprocess \
--noconcat_box_obj_quality_cls
```

</details>

#### 1-2. (Optional) Provide anchors/wh_scale as .npy

<details><summary>Click to expand</summary>

If `*_anchors.npy` and `*_wh_scale.npy` exist, QAT will load them automatically.
It searches the ONNX directory first, then the current working directory.

```
ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost_nocat_anchors.npy
ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost_nocat_wh_scale.npy
```

If they are missing, QAT falls back to anchors/wh_scale stored inside the ONNX.

</details>

#### 1-3. Run QAT

<details><summary>Click to expand</summary>

```bash
uv run python uhd/uhd_qat.py \
--onnx-model ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_static_nopost_nocat.onnx \
--image-dir data/wholebody34/obj_train_data \
--img-size 64x64 \
--resize-mode opencv_inter_nearest \
--class-ids 0 \
--batch-size 64 \
--epochs 10 \
--target esp32s3 \
--num-of-bits 8 \
--lr 1e-5
```

- Outputs are saved under `runs/uhd_qat/epoch_***` and `runs/uhd_qat/best_***`
as `.espdl` and `.native`.
- `--resize-mode` must match the training/export setting (e.g., `opencv_inter_nearest`).
- If the model has a quality head, QAT usually detects it automatically.
  If unsure, force it with `--use-quality` or `--no-quality`.
- Layerwise equalization is **enabled by default** with the same settings as
  `uhd/quantize_onnx_model_for_esp32.py`. Disable via `--no-equalization`.

</details>

#### 1-4. CLI options (QAT)

<details><summary>Click to expand</summary>

- Required: `--onnx-model`, `--image-dir`
- Data: `--img-size`, `--resize-mode`, `--class-ids`, `--list-path`, `--val-split`, `--batch-size`, `--num-workers`
- Quant: `--target`, `--num-of-bits`, `--calib-steps`
- Equalization: `--no-equalization`, `--equalization-iterations`, `--equalization-value-threshold`, `--equalization-opt-level`
- Anchor head: `--use-quality`, `--no-quality`, `--iou-loss`, `--anchor-assigner`, `--anchor-cls-loss`, `--simota-topk`
- Training: `--epochs`, `--lr`, `--momentum`, `--weight-decay`, `--seed`, `--device`
- Eval/output: `--no-eval`, `--conf-thresh`, `--nms-thresh`, `--iou-thresh`, `--score-mode`, `--quality-power`, `--save-dir`, `--output-prefix`

</details>

### 2. PTQ (Post-Training Quantization)

This repository includes a calibration/quantization script for ESP-DL:
`uhd/quantize_onnx_model_for_esp32.py`.

#### 2-1. Image-only calibration

<details><summary>Click to expand</summary>

```bash
### Example command that prioritizes maximum speed at the expense of accuracy
uv run python uhd/quantize_onnx_model_for_esp32.py \
--dataset-type image \
--image-dir data/wholebody34/obj_train_data \
--resize-mode opencv_inter_nearest \
--onnx-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx \
--espdl-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.espdl \
--target esp32s3 \
--calib-algorithm kl \
--use-layerwise-equalization

### Example command to minimize quantization error
uv run python uhd/quantize_onnx_model_for_esp32.py \
--dataset-type image \
--image-dir data/wholebody34/obj_train_data \
--resize-mode opencv_inter_nearest \
--onnx-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx \
--espdl-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.espdl \
--target esp32s3 \
--calib-algorithm kl \
--use-layerwise-equalization \
--int16-op-pattern /box_out/Conv \
--int16-op-pattern /quality_out/Conv \
--int16-op-pattern /backbone/block1/pw/conv/Conv \
--int16-op-pattern /backbone/block1/pw/act/Relu

uv run python uhd/quantize_onnx_model_for_esp32.py \
--dataset-type image \
--image-dir data/wholebody34/obj_train_data \
--resize-mode opencv_inter_nearest \
--onnx-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.onnx \
--espdl-model ultratinyod_anc8_w32_64x64_opencv_inter_nearest_static_nopost.espdl \
--target esp32s3 \
--calib-algorithm kl \
--use-layerwise-equalization \
--int16-op-pattern /box_out/Conv \
--int16-op-pattern /quality_out/Conv \
--int16-op-pattern /box_tower/box_tower.1/pw/conv/Conv \
--int16-op-pattern /box_tower/box_tower.1/pw/act/Relu \
--int16-op-pattern /depthwiseconv/backbone/block1/dw/conv/Conv \
--int16-op-pattern /backbone/block1/dw/act/Relu \
--int16-op-pattern /backbone/stem/conv/Conv \
--int16-op-pattern /backbone/stem/act/Relu \
--int16-op-pattern /backbone/block1/pw/conv/Conv \
--int16-op-pattern /backbone/block1/pw/act/Relu \
--int16-op-pattern /backbone/block2/pw/conv/Conv \
--int16-op-pattern /backbone/block2/pw/act/Relu \
--int16-op-pattern /depthwiseconv/backbone/block2/dw/conv/Conv \
--int16-op-pattern /backbone/block2/dw/act/Relu \
--int16-op-pattern /backbone/block3_skip/conv/Conv \
--int16-op-pattern /backbone/sppf/scale_x/conv/Conv
##################################################################

Notes:
- The YUV422 models expect input shape `[1, 2, 64, 64]` with `opencv_inter_nearest_yuv422` preprocessing.
- `--dataset-type image` is the default and ignores labels.
- Adjust `--calib-steps`, `--batch-size`, `--target`, `--num-of-bits`, and `--device` as needed.

</details>

#### 2-2. CLI options

<details><summary>Click to expand</summary>

- `--image-dir`: Directory containing calibration images.
- `--dataset-type`: Calibration dataset type (`image` or `yolo`, default `image`).
- `--list-path`: Optional text file listing images to use.
- `--export-anchors-wh-scale-dir`: Directory to save `{onnx-model}_anchors.npy` and `{onnx-model}_wh_scale.npy` (default: same directory as `--espdl-model`).
- `--expand-group-conv`: Expand `groups > 1` conv into group=1 (default: disabled).
- `--img-size`: Square input size used for calibration (default `64`).
- `--resize-mode`: Resize mode (default `opencv_inter_nearest_yuv422`).
- `--class-ids`: Comma-separated class IDs to keep (yolo only, default `0`).
- `--split`: Dataset split for calibration (`train`, `val`, `all`, default `all`).
- `--val-split`: Validation split ratio (ignored when `--split all`, default `0.0`).
- `--batch-size`: Calibration batch size (default `1`).
- `--calib-steps`: Number of calibration steps (default `32`).
- `--calib-algorithm`: Calibration algorithm (default `kl`; examples: `minmax`, `mse`, `percentile`).
- `--use-layerwise-equalization`: Enable layerwise equalization quantization (default: disabled).
- `--int16-op-pattern`: Regex pattern to force matched ops to int16 (repeatable).
- `--onnx-model`: Path to the input ONNX model (default `ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.onnx`).
- `--espdl-model`: Path to the output `.espdl` file (default `ultratinyod_res_anc8_w16_64x64_opencv_inter_nearest_yuv422_distill_static_nopost.espdl`).
- `--target`: Quantize target type (`c`, `esp32s3`, `esp32p4`, default `esp32s3`).
- `--num-of-bits`: Quantization bits (default `8`).
- `--device`: Device for calibration (`cpu` or `cuda`, default `cpu`).

</details>

## Ultra-lightweight classification model series

1. [VSDLM: Visual-only speech detection driven by lip movements](https://github.com/PINTO0309/VSDLM) - MIT License
2. [OCEC: Open closed eyes classification. Ultra-fast wink and blink estimation model](https://github.com/PINTO0309/OCEC) - MIT License
3. [PGC: Ultrafast pointing gesture classification](https://github.com/PINTO0309/PGC) - MIT License
4. [SC: Ultrafast sitting classification](https://github.com/PINTO0309/SC) - MIT License
5. [PUC: Phone Usage Classifier is a three-class image classification pipeline for understanding how people
interact with smartphones](https://github.com/PINTO0309/PUC) - MIT License
6. [HSC: Happy smile classifier](https://github.com/PINTO0309/HSC) - MIT License
7. [WHC: Waving Hand Classification](https://github.com/PINTO0309/WHC) - MIT License
8. [UHD: Ultra-lightweight human detection](https://github.com/PINTO0309/UHD) - MIT License
9. [MWC: Mask wearing classifier](https://github.com/PINTO0309/MWC) - MIT License
10. [SGC: Classification of wearing vs. not wearing sunglasses. 48x48.](https://github.com/PINTO0309/SGC) - MIT License
11. [HHC: Head Hat Classification. HHC is a binary classifier for cropped head images. 48x48.](https://github.com/PINTO0309/HHC) - MIT License
12. [BPC: Background Plain classification. 48x48.](https://github.com/PINTO0309/BPC) - MIT License
13. [PPC: Binary classification to determine whether the subject is holding a smartphone. 48x48 RGB image.](https://github.com/PINTO0309/PPC) - MIT License

## Citation

If you find this project useful, please consider citing:

```bibtex
@software{hyodo2025uhd,
  author    = {Katsuya Hyodo},
  title     = {PINTO0309/UHD},
  month     = {12},
  year      = {2025},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.17790207},
  url       = {https://github.com/PINTO0309/uhd},
  abstract  = {Ultra-lightweight human detection. The number of parameters does not correlate to inference speed.},
}
```

## Ref

- https://github.com/espressif/esp-dl
- https://github.com/espressif/esp-who
- https://github.com/espressif/esp-who/blob/master/docs/en/get-started/ESP32-S3-EYE_Getting_Started_Guide.md
- [ESP32 - ESP-IDF](https://zenn.dev/pinto0309/scraps/6b985872ea9a89)
  - ESP-WHO + human_face_recognition

https://github.com/user-attachments/assets/f67c0470-f74a-4a70-9e14-213bfc141064

- ESP32 UHD: https://github.com/PINTO0309/esp-who/tree/custom/examples/ultra_lightweight_human_detection
  - ESP32-S3-EYE - Emphasis on precision `R`

https://github.com/user-attachments/assets/ad417641-0c06-4307-9224-a1bdb9402dbd

UHD

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README