SR-TOD

Introduce

This is the official code for the paper Visible and Clear: Finding Tiny Objects in Difference Map.

The link to the paper is https://arxiv.org/abs/2405.11276.

This project is built based on mmdetection 3.1 and mmcv 2.0.1.

NOTE: Our paper has been accepted by ECCV 2024.

Environment

pytorch 1.12.0

torchvision 0.13.0

mmdetection 3.1

mmcv 2.0.1

The installation and usage of mmdetection can be referred to at the following link: https://mmdetection.readthedocs.io/en/latest/get_started.html.

To use the AI-TOD evaluation metrics, you need to download aitodpycocotools. You can install it using the following command:

pip install "git+https://github.com/jwwangchn/cocoapi-aitod.git#subdirectory=aitodpycocotools"

For other environment requirements, please refer to mmdetection.

Training and Test

The training and test commands can also be referenced from mmdetection.

1 gpu:

python tools/train.py ./srtod_project/srtod_cascade_rcnn/config/srtod-cascade-rcnn_r50_fpn_1x_coco.py

python tools/test.py ./srtod_project/srtod_cascade_rcnn/config/srtod-cascade-rcnn_r50_fpn_1x_coco.py your_model.pth

If you need to use more GPUs, you should use ./tools/dist_train.sh instead of tools/train.py.

DroneSwarms

If you want to access the DroneSwarms dataset, please visit the following link：DroneSwarms

Reference

@article{cao2024visible,
  title={Visible and Clear: Finding Tiny Objects in Difference Map},
  author={Cao, Bing and Yao, Haiyu and Zhu, Pengfei and Hu, Qinghua},
  journal={arXiv preprint arXiv:2405.11276},
  year={2024}
}