Rethinking Label Assignment and Sampling Strategies for Two-Stage Oriented Object Detection

note: We mainly disclose the code based on RPGAOD here. (i.e. RPGAOD+OAS)

Introduction

This article proposes a new direction aware allocator and sampler (OAS) to alleviate the problem of angle periodicity by improving label allocation and sampling strategies. It combines angle and position information to optimize the allocation of positive and negative samples, and adjusts learning weights based on angle differences, thereby significantly improving the performance of the two-stage rotating object detector. The experimental results indicate that OAS can achieve higher detection accuracy on multiple challenging rotating target detection benchmarks without increasing additional costs.

Motivation

Overall method

Get Started

(1) Prepare environment

# create environment
conda create -n RPGAOD_oas python=3.7 pip --yes
conda activate RPGAOD_oas

# download and install pytorch and torchvision
wget https://download.pytorch.org/whl/cu113/torch-1.10.0%2Bcu113-cp37-cp37m-linux_x86_64.whl#sha256=2a90db92579ed875d2aa0ad6af55359e3f0426a523056a7b49b002c3cc6d2ad8
wget https://download.pytorch.org/whl/cu113/torchaudio-0.10.0%2Bcu113-cp37-cp37m-linux_x86_64.whl#sha256=16ca574d33837643b49cf5097099bcfaf2c8bdc9146bc8117c13c14fd2dc28fc
wget https://download.pytorch.org/whl/cu113/torchvision-0.11.0%2Bcu113-cp37-cp37m-linux_x86_64.whl#sha256=d0cbfedc7ee6ede94d1f60531fc178691e9cf1cddedcb5d7168fd0dc8249e46e
pip install torch-1.10.0+cu113-cp37-cp37m-linux_x86_64.whl
pip install torchaudio-0.10.0+cu113-cp37-cp37m-linux_x86_64.whl
pip install torchvision-0.11.0+cu113-cp37-cp37m-linux_x86_64.whl

# clone this codebase, including submodule
git clone https://github.com/skyandkibo/OAS.git --recursive
cd RPGAOD

# install submodule
cd BboxToolkit
pip install -v -e .
cd ..

cd mmdet
cd ops
cd orn
pip install -v -e .
cd ../../..

# install mmcv
pip install mmcv-full==1.3.17 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.10.0/index.html

# install other requirements
pip install mmpycocotools
pip install timm==0.8.0dev0
pip install yapf==0.40.0

# install this codebase
pip install -v -e .

(2) Prepare data

Download the DOTA dataset from official website, change the diretory in ./BboxToolkit/tools/split_configs/dota1_0/ss_trainval.json and ./BboxToolkit/tools/split_configs/dota1_0/ss_test.json, and then run the following script

cd ./BboxToolkit/tools/
python img_split.py --base_json split_configs/dota1_0/ss_trainval.json
python img_split.py --base_json split_configs/dota1_0/ss_test.json

Download the DIOR-R dataset from official website

Usage

Training

DOTA

Single GPU
python tools/train.py configs/obb/RPGAOD/SR2_ORCNN_r50_dota10_oas.py

Multi-GPU
./tools/dist_train.py configs/obb/RPGAOD/SR2_ORCNN_r50_dota10_oas.py 4 29500

DIOR

Single GPU
python tools/train.py configs/obb/RPGAOD/SR2_ORCNN_r50_dior10_oas.py

Multi-GPU
./tools/dist_train.py configs/obb/RPGAOD/SR2_ORCNN_r50_dior10_oas.py 4 29500

Testing

DOTA

python tools/test.py configs/obb/RPGAOD/SR2_ORCNN_r50_dota10_oas.py work_dirs/SR2_ORCNN_r50_dota10_oas/epoch_36.pth --format-only --options save_dir=保存文件夹名字

DIOR

python tools/test.py configs/obb/RPGAOD/SR2_ORCNN_r50_dior10_oas.py work_dirs/SR2_ORCNN_r50_dior10_oas/epoch_36.pth --eval mAP

We provide the pretrained models as listed below. (Trained on 4 RTX 3090 GPUs, tested on 1 RTX 3090 GPU)

Detector	Dataset	box AP (in paper)	box AP (this repo)	pretrained model & config
RPGAOD	DOTA	76.47	76.50	BaiduCloud
RPGAOD+OAS	DOTA	77.90	78.11	BaiduCloud
RPGAOD	DIOR-R	63.87	63.85	BaiduCloud
RPGAOD+OAS	DIOR-R	64.41	64.57	BaiduCloud

Acknowledgement

This code is developed on the top of RPGAOD, we thank to their efficient and neat codebase.

Citation

If you find our work is useful in your research, please consider citing:

@article{qian2025rethinking,
  title={Rethinking Label Assignment and Sampling Strategies for Two-Stage Oriented Object Detection},
  author={Qian, Jinjin and Liu, Chengxu and Ai, Yubin and Zhao, Guoshuai and Qian, Xueming},
  journal={IEEE Geoscience and Remote Sensing Letters},
  year={2025},
  publisher={IEEE}
}