HazyDet: Open-Source Benchmark for Drone-View Object Detection With Depth-Cues in Hazy Scenes

This repository is the official implementation of HazyDet

• HazyDet
  • Leadboard and Model Zoo
    • Detectors
    • Dehazing
• DeCoDet
  • Installation
    • Step 1: Create a conda
    • Step 2: Install PyTorch
    • Step 3: Install OpenMMLab 2.x Codebases
    • Step 4: Install HazyDet
  • Training
  • Inference
  • Depth Maps
• Acknowledgement
• Citation

HazyDet

You can download our HazyDet dataset from Baidu Netdisk or OneDrive.

For both training and inference, the following dataset structure is required:

HazyDet |-- train |-- clean images |-- hazy images |-- labels |-- val |-- clean images |-- hazy images |-- labels |-- test |-- clean images |-- hazy images |-- labels |-- RDDTS |-- hazy images |-- labels

Note: Both passwords for BaiduYun and OneDrive is grok.

Leadboard and Model Zoo

All the weight files in the model zoo can be accessed on Baidu Cloud and OneDrive.

Detectors

Dehazing

DeCoDet

Installation

Step 1: Create a conda

shell $ conda create --name HazyDet python=3.9 $ source activate HazyDet

Step 2: Install PyTorch

shell conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Step 3: Install OpenMMLab 2.x Codebases

```shell

openmmlab codebases

pip install -U openmim --no-input mim install mmengine "mmcv>=2.0.0" "mmdet>=3.0.0" "mmsegmentation>=1.0.0" "mmrotate>=1.0.0rc1" mmyolo "mmpretrain>=1.0.0rc7" 'mmagic'

other dependencies

pip install -U ninja scikit-image --no-input ```

Step 4: Install HazyDet

shell python setup.py develop

Note: make sure you have cd to the root directory of HazyDet

shell $ git clone git@github.com:GrokCV/HazyDet.git $ cd HazyDet

Training

shell $ python tools/train_det.py configs/DeCoDet/DeCoDet_r50_1x_hazydet.py

Inference

shell $ python tools/test.py configs/DeCoDet/DeCoDet_r50_1x_hazydet.py weights/fcos_DeCoDet_r50_1x_hazydet.pth

We released our checkpoint on HazyDet

Depth Maps

The depth map required for training can be obtained through Metic3D. They can also be acquired through other depth estimation models.

If you want to use our depth data, please download it and place it in the specified path. For convenience in storage and viewing, we save relative depth in PNG image format and the maximum depth in text format, but we use absolute depth during training.

Acknowledgement

We are grateful to the Tianjin Key Laboratory of Visual Computing and Intelligent Perception (VCIP) for providing essential resources. Our sincere appreciation goes to Professor Pengfei Zhu and the dedicated AISKYEYE team at Tianjin University for their invaluable support with data, which has been crucial to our research efforts. We also deeply thank Xianghui Li, Yuxin Feng, and other researchers for granting us access to their datasets, significantly advancing and promoting our work in this field. Additionally, our thanks extend to Metric3D for its contributions to the methodology presented in this article.

Citation

If you use this toolbox or benchmark in your research, please cite this project.

```bibtex @article{feng2024HazyDet, title={HazyDet: Open-source Benchmark for Drone-view Object Detection with Depth-cues in Hazy Scenes}, author={Feng, Changfeng and Chen, Zhenyuan and Kou, Renke and Gao, Guangwei and Wang, Chunping and Li, Xiang and Shu, Xiangbo and Dai, Yimian and Fu, Qiang and Yang, Jian}, year={2024}, journal={arXiv preprint arXiv:2409.19833}, }

@article{zhu2021detection, title={Detection and tracking meet drones challenge}, author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Fan, Heng and Hu, Qinghua and Ling, Haibin}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, volume={44}, number={11}, pages={7380--7399}, year={2021}, publisher={IEEE} } ```

test

test