Hausdorff Distance Matching with Adaptive Query Denoising for Rotated Detection Transformer

SI Analytics

Hakjin Lee, Minki Song, Jamyoung Koo, Junghoon Seo

[Paper]

The RHINO is a robust DETR architecture designed for detecting rotated objects. It demonstrates promising results, exceeding 60 mAP on DOTA-v2.0.

Main Results

DOTA-v2.0 (Single-Scale Training and Testing)

Method	Backbone	AP50	Config	Download
RHINO	ResNet50 (1024,1024,200)	59.26	rhino_r50_dota2	model
RHINO	Swin-T (1024,1024,200)	60.72	rhino_swint_dota2	model

DOTA-v1.5 (Single-Scale Training and Testing)

Method	Backbone	AP50	Config	Download
RHINO	ResNet50 (1024,1024,200)	71.96	rhino_r50_dotav15	model
RHINO	Swin-T (1024,1024,200)	73.46	rhino_swint_dotav15	model

DOTA-v1.0 (Single-Scale Training and Testing)

Method	Backbone	AP50	Config	Download
RHINO	ResNet50 (1024,1024,200)	78.68	rhino_r50_dota	model
RHINO	Swin-T (1024,1024,200)	79.42	rhino_swint_dota	model

Requirements

Installation

# torch>=1.9.1 is required.
pip install openmim mmengine==0.7.3
mim install mmcv==2.0.0
pip install mmdet==3.0.0
pip3 install --no-cache-dir -e ".[optional]"

or check the Dockerfile.

Preprare Dataset

Details are described in https://github.com/open-mmlab/mmrotate/blob/main/tools/data/dota/README.md

Specifically, run below code.

python3 tools/data/dota/split/img_split.py --base-json \
  tools/data/dota/split/split_configs/ss_trainval.json

python3 tools/data/dota/split/img_split.py --base-json \
  tools/data/dota/split/split_configs/ss_test.json

Training

To train the model(s) in the paper, run this command:

# DOTA-v2.0 R-50
export CONFIG='configs/rhino/rhino_phc_haus-4scale_r50_2xb2-36e_dotav2.py'
bash tools/dist_train.sh $CONFIG 2

Evaluation

To evaluate our models on DOTA, run:

# example
export CONFIG='configs/rhino/rhino_phc_haus-4scale_r50_2xb2-36e_dotav2.py'
export CKPT='work_dirs/rhino_phc_haus-4scale_r50_2xb2-36e_dotav2/epoch_36.pth'
python3 tools/test.py $CONFIG $CKPT

Evaluation is processed in the official DOTA evaluation server.

License

This project is licensed under CC-BY-NC. It is available for academic purposes only.

Bibtex

If you find our work helpful for your research, please consider citing the following BibTeX entry.

@misc{lee2023hausdorff,
      title={Hausdorff Distance Matching with Adaptive Query Denoising for Rotated Detection Transformer},
      author={Hakjin Lee and Minki Song and Jamyoung Koo and Junghoon Seo},
      year={2023},
      eprint={2305.07598},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}