WeaveEarth

KO | EN

gitlite — search

#python #typescript #golang #android #rust #nodejs #macos #java #awesome-list #vue #security #pytorch #ios #open-source

(★ 15)

No description available.

WeaveEarth Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">
 <img src="assets/logo.png" alt="WeaveEarth Logo" width="39" style="vertical-align: middle; margin-right: 0px;">
 WeaveEarth: Structured Evidence Construction and Reasoning 
 for Training-Free UHR Remote Sensing Understanding
</h1>

Xianzhi Ma1, Shujun Wang1, Xiaohan Li1, Hao Liu1, Changhua Pei2, Jianhui Li1,*

1 Institute of Space Earth Science, School of Frontier Sciences, Nanjing University, Suzhou, China 
 2 Computer Network Information Center, Chinese Academy of Sciences, Beijing, China

---

> **Abstract:** *Ultra-High-Resolution (UHR) remote sensing image understanding requires Vision-Language Models (VLMs) to capture both the global scene layout and sparse yet task-critical local details under limited computational budgets. Existing methods mainly follow two paradigms. One is passive perception, which relies on resolution expansion or token compression and may therefore discard fine-grained details. The other is active perception, which depends on multi-round zooming and search, but suffers from high latency, contextual fragmentation, and error accumulation. We argue that a more effective path toward UHR understanding lies not in accessing more, but in organizing better. To this end, we propose WeaveEarth, a training-free framework that reformulates UHR understanding as a problem of structured evidence construction and reasoning under global context constraints. Specifically, WeaveEarth first employs Global-Aware Evidence Construction to select a compact, low-redundancy, and spatially complementary Minimal Support Evidence Set. It then introduces Structured Evidence Reasoning, which weaves local evidence, spatial metadata, and relative topology into a unified reasoning interface, thereby enhancing the VLM's ability to perform global-local joint reasoning. Extensive experiments show that WeaveEarth consistently outperforms strong baselines and existing UHR methods across multiple UHR remote sensing benchmarks and multiple frozen VLM backbones.*

</div>

## 📑 Contents

- [🔥 News](#news)
- [🛠️ Setup](#setup)
- [💿 Benchmarks](#benchmarks)
- [🚀 Models](#models)
- [🏋 Evaluation](#evaluation)

---

<a id="news"></a>
## 🔥 News

- **[2026-07]**: 🎉 WeaveEarth is accepted by **ACM Multimedia 2026**!
- **[2026-04]**: ⏳ We have released the initial evaluation script for **Qwen3-VL** on the **LRS-VQA** benchmark. Other evaluation scripts are coming soon.

<a id="setup"></a>
## 🛠️ Setup

### Environment

- Python 3.11.11
- PyTorch 2.5.1
- NVIDIA A100 80GB GPU
- CUDA 12.4

### Installation

```bash
cd /WeaveEarth
conda create -n WeaveEarth python=3.11.11 -y
conda activate WeaveEarth
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt
```

<a id="benchmarks"></a>
## 💿 Benchmarks

We evaluate WeaveEarth on three representative UHR RS image understanding benchmarks. Please place the downloaded benchmarks in `/WeaveEarth/datars`.

<a id="models"></a>
## 🚀 Models

Our method is a training-free framework that can be instantiated on representative open-source VLM backbones without any parameter fine-tuning. Please place the downloaded pre-trained weights in the `/WeaveEarth/ckpts` directory. In addition to the reasoning backbone, please also download the semantic retrieval encoder [`siglip2-so400m-patch16-naflex`](https://huggingface.co/google/siglip2-so400m-patch16-naflex-jax) and place it in the `/WeaveEarth/ckpts` directory.

| Backbone | Parameters | Paradigm | LRS-VQA (%) | MME-RealWorld (%) | XLRS-Bench (%) |
| :-- | :--: | :--: | --: | --: | --: |
| [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL) | 8B | Training-Free | 33.38 | 47.38 | 47.14 |
| [LLaVA-v1.6](https://github.com/haotian-liu/llava) | 7B | Training-Free | 27.60 | 28.04 | 28.87 |
| [IXC-2.5](https://github.com/InternLM/InternLM-XComposer) | 7B | Training-Free | 32.58 | 39.96 | 34.37 |

<a id="evaluation"></a>
## 🏋 Evaluation
Please run the following command for evaluation:
```bash
python /WeaveEarth/inference_scripts/eval_lrs_vqa_qwen3vl.py --reasoning_model_path /WeaveEarth/ckpts/Qwen3-VL-8B-Instruct --dataset_jsonl_path /WeaveEarth/datars/LRS-VQA/LRS_VQA_merged.jsonl --dataset_root /WeaveEarth/datars/LRS-VQA --save_path /WeaveEarth/output/LRS-VQA-WeaveEarth
```
After evaluation, please check the specified `save_path` directory (e.g., `/WeaveEarth/output/LRS-VQA-WeaveEarth`) for the saved predictions and summary results.

WeaveEarth

WeaveEarth Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

WeaveEarth

WeaveEarth Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README