DAFNe: A One-Stage Anchor-Free Approach for Oriented Object Detection

<img src="./res/header.png"/>

Code for our Paper DAFNe: A One-Stage Anchor-Free Approach for Oriented Object Detection.

</br> </br>

Datasets

• UCAS-AOD: https://hyper.ai/datasets/5419
• DOTA 1.0/1.5: https://captain-whu.github.io/DOTA/index.html
  • Note: See ./tools/prepare_dota/ for instructions on how to prepare the DOTA datasets.
• HRSC2016: https://www.kaggle.com/guofeng/hrsc2016

Docker Setup

Use the Dockerfile to build the necessary docker image:

docker build -t dafne .

Training

Check out ./configs/pre-trained/ for different pre-defined configurations for the DOTA 1.0, DOTA 1.5, UCAS-AOD, and HRSC2016 datasets. Use these paths as argument for the --config-file option below.

With Docker

Use the ./tools/run.py helper to start running experiments

./tools/run.py --gpus 0,1,2,3 --config-file ./configs/dota-1.0/1024.yaml

Without Docker

NVIDIA_VISIBLE_DEVICES=0,1,2,3 ./tools/plain_train_net.py --num-gpus 4 --config-file ./configs/dota-1.0/1024.yaml

Pre-Trained Weights

| Dataset | mAP (%) | Config | Weights | |----------|---------|-----------------------------------------------------------------|------------------------------------------------------------------------------------------------------------| | UCAS-AOD | 89.65 | ucas_aod_r101_ms | ucas-aod-r101-ms.pth | | HRSC2016 | 89.76 | hrsc_r50_ms | hrsc-r50-ms.pth | | DOTA 1.0 | 76.95 | dota-1.0_r101_ms | dota-1.0-r101-ms.pth | | DOTA 1.5 | 71.99 | dota-1.5_r101_ms | dota-1.5-r101-ms.pth |

Pre-Trained Weights Usage with Docker

./tools/run.py --gpus 0 --config-file <CONFIG_PATH> --opts "MODEL.WEIGHTS <WEIGHTS_PATH>"

Pre-Trained Weights Usage without Docker

NVIDIA_VISIBLE_DEVICES=0 ./tools/plain_train_net.py --num-gpus 1 --config-file <CONFIG_PATH> MODEL.WEIGHTS <WEIGHTS_PATH>

Cite

@misc{lang2021dafne,
      title={DAFNe: A One-Stage Anchor-Free Approach for Oriented Object Detection}, 
      author={Steven Lang and Fabrizio Ventola and Kristian Kersting},
      year={2021},
      eprint={2109.06148},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgments

• Thanks to AdelaiDet for providing the initial FCOS implementation
• Thanks to Detectron2 for providing a general object detection framework