Create and activate a conda environment: conda create -n e2e-osdet python=3.11 conda activate e2e-osdet Install the required packages: git clone https://github.com/wchao0601/M4-SAR.git cd M4-SAR/ pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118 pip install seaborn thop timm einops cd STTrack/mamba_install/causal-conv1d CAUSAL_CONV1D_FORCE_BUILD=TRUE pip install . cd ../selective_scan pip install . cd M4-SAR/ pip install -r requirements.txt We provide a complete environment configuration log for your reference.

Optical and SAR Attributes by File Range

| File Range | Optical Image Resolution | SAR Image Polarization | |:---:|:---:|:---:| | 1.jpg ~ 56087.jpg | 10 meters | VH | | 56088.jpg ~ 112174.jpg | 60 meters | VV |

Data Preparation

| Dataset | Link1 | Link2 | Link3 | SR & Pola. | Image Size | Category | Ins.num | Img.num | | :---: | :---: | :---: | :---: | :---:| :---: | :---: | :---: | :---: | | M4-SAR | Kaggle|Baidu|Hug-Face|10M, 60M, VH, VV|512 x 512|6|981,862|112,174|

Dataset and Label Structure

Single-GPU Train

```python

please set 'device=0' in train.py

python train.py ```

Multi-GPU Train

```python

please set 'device=[0,1]' in train.py

python multigpu-train.py ```

Test

python python test.py

Gen-Predict

python python gen-predict-label.py

Vis-Predict

python python vis-predict-label.py

Gen-Heatmap

python python gen-heatmap.py

Results

| Model |Weight Link | Img size (pixels) | #Para(M) | Tra.Time (h) | Inf.Time (ms) | AP50 (%) | AP75 (%) | mAP (%) | | :---: | :---: | :---:| :---: | :---: | :---: | :---: | :---: | :---: | | CFT |Download | 512 x 512 | 53.8 | 60.6 | 40.6 | 84.6 | 68.9 | 59.9 | | CLANet |Download | 512 x 512 | 48.2 | 56.2 | 29.1 | 84.6 | 68.5 | 59.6 | | CSSA |Download | 512 x 512 | 13.5 | 25.7 | 12.3 | 83.4 | 66.4 | 58.0 | | CMADet |Download | 512 x 512 | 41.5 | 52.4 | 46.7 | 81.5 | 63.5 | 55.7 | | ICAFusion |Download | 512 x 512 | 29.0 | 47.7 | 23.6 | 84.5 | 67.3 | 58.8 | | MMIDet |Download | 512 x 512 | 53.8 | 49.9 | 41.9 | 84.8 | 68.6 | 59.8 | | E2E-OSDet |Download | 512 x 512 | 27.5 | 42.1 | 20.9 | 85.7 | 70.3 | 61.4 |

All weights

Google Drive

Contact

If you have any questions, please feel free to contact me via email at wchao0601@163.com

Citation

If our work is helpful, you can cite our paper: ``` @article{wang2025m4, title={M4-SAR: A Multi-Resolution, Multi-Polarization, Multi-Scene, Multi-Source Dataset and Benchmark for Optical-SAR Fusion Object Detection}, author={Wang, Chao and Lu, Wei and Li, Xiang and Yang, Jian and Luo, Lei}, journal={arXiv preprint arXiv:2505.10931}, year={2025} }

```

Acknowledgment

• This repo is based on Ultralytics, CFT, CLANet, CSSA, CMADet, ICAFusion and MMIDet which are excellent works.
• We thank the STTrack and YOLOv12 libraries, which help us to implement our ideas quickly.