ElDet

We propose an anchor-free general ellipse object detector that can better detect ellipse objects of any class based on object shape information, and can be applied to downstream tasks such as face detection with a few modifications.

1. Environment Configuration

python>=3.6
torch>=1.7.0
others see requirements.txt

Use the requirements.txt to build the basic environment;
DCNv2
```
cd DCNv2
sh make.sh
```
Copy the folder ./DCNv2/build to ./dcn.

2. Data Format

2.1 Data annotation

We use VIA to make the labels, and export .json file. The transform the JSON format to COCO format.

data_process.py is a routine for format transformation.

2.2 Data format

We adapt COCO format and bbox = [cx, cy, a, b, θ], where cx, cy are center point coordinates, a, b are major axis and minor axis of ellipse, θ ∈[-90, 90） is rotation angle of ellipse.

Data folder format

--data
  --your data name
    --images
      --1.jpg
      --2.jpg
      ...    
    --annotations
      --train.json
      --test.json

3. Detection Results

3.1. GED dataset

GED dataset download link (Baidu Netdisk)

3.2. FDDB dataset

4. Citataion

@inproceedings{liao2020speech2video,
  title={ElDet: An Anchor-free General Ellipse Object Detector},
  author={Wang, Tian hao and Lu, Changsheng and Shao, Ming and Yuan, Xiaohui and Xia, Siyu},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2022}
}

@article{lu2019arc,
  title={Arc-Support Line Segments Revisited: An Efficient High-Quality Ellipse Detection},
  author={Lu, Changsheng and Xia, Siyu and Shao, Ming and Fu, Yun},
  journal={IEEE Transactions on Image Processing},
  volume={29},
  pages={768--781},
  year={2020},
  publisher={IEEE}
}

One1h/ElDet