RangeCM

KO | EN

gitlite — search

#javascript #ai #deep-learning #linux #docker #nodejs #windows #awesome #cli #vue #awesome-list #ios #mcp #open-source

(★ 14)

Low-Latency Neural LiDAR Compression with 2D Context Models

RangeCM 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Low-Latency Neural LiDAR Compression with 2D Context Models

## Introduction
This repository presents the official code to establish RangeCM, proposed in the paper **Low-Latency Neural LiDAR Compression with 2D Context Models** (ICLR 2026).

![RangeCM_pipeline](https://raw.githubusercontent.com/rrui-song/RangeCM/master/assets/RangeCM.png)

## Setup
### 1. Build Environment
Please first check the `environment.sh` file and modify the CUDA version for pytorch installation according to your device. Then, install the dependencies using:

```
conda create -n RangeCM python=3.10 -y
conda activate RangeCM
bash environment.sh
```

### 2. Prepare Dataset
Please first download the Waymo Open Dataset and the KITTI Dataset.

* [Waymo Open Dataset v1.4.0](https://waymo.com/intl/es/open/download#perception-dataset)

* [KITTI odometry](https://www.cvlibs.net/datasets/kitti/eval_odometry.php)

Then, process the files following the instructions in `data/prepare_dataset.md`.

### 3. Evaluation
First, please download the pretrained models from the [huggingface repository](https://huggingface.co/rrui-song/RangeCM/tree/main).

Next, run `geometry_coder.py` and `geometry_intensity_coder.py` to evaluate geometry and joint geometry-intensity compression, respectively. For example, the following code performs geometry-intensity compression on the Waymo Open Dataset.

```
python geometry_intensity_coder.py 
    --config_file "./config/eval/waymo_geometry_intensity.yaml"
    --data_path "/dataset/waymo/testing/range_image/" 
    --binfile_path "./output/waymo_geometry_intensity"
    --device 0
```

The compression configurations can be adjusted by modifying the `yaml` files in the `config/eval` folder.

### 4. Training
Training scripts are provided in the `train` folder, and the configurations can be adjusted by the `yaml` files in the `config/train` folder. The scripts support distributed data parallel (DDP) training. To enable DDP, please set the `enable_DDP` variable in configuration files to `True`. Then, the DDP training can be launched as following:

```
torchrun --nproc_per_node 2 \
    train/train_geometry_intensity.py \
    --config_file "./config/train/KITTI_geometry_intensity.yaml" \
    --data_path "/dataset/KITTI/testing/range_image/" 
```

## Citation
If this repository is helpful for your research, please consider citing our paper:

```
@inproceedings{
song2026lowlatency,
title={{Low-Latency Neural LiDAR Compression with 2D Context Models}},
author={Song, Rui and Wang, Yan and Xu, Tongda and Liu, Zhening and Lin, Zehong and Zhang, Jun},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}
```

RangeCM

RangeCM 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

RangeCM

RangeCM 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README