Proxy Anchor-based Unsupervised Learning for Continuous Generalized Category Discovery

Official PyTorch implementation of ICCV 2023 paper

Proxy Anchor-based Unsupervised Learning for Continuous Generalized Category Discovery.

Requirements

• Python3
• PyTorch (> 1.0)
• NumPy
• tqdm

Datasets

1. Download four public benchmarks for fine-grained dataset

   • CUB-200-2011
   • MIT-67: Indoor Scene Recognition
   • Stanford Dogs Dataset
   • FGVC-Aircraft Benchmark

2. Extract the tgz or zip file into ./data/ (Exceptionally, for CUB-200-2011, put the files in a ./data/CUB200)

Train Examples

• CUB-200: We used 1 GPUs to train CUB-200.

  python train.py \
  --model resnet18 \
  --dataset cub \
  --alpha 32 \
  --mrg 0.1 \
  --lr 1e-4 \
  --warm 5 \
  --epochs 60 \
  --batch_size 120 \

Acknowledgements

Our code is modified and adapted on these great repositories:

• Proxy Anchor Loss for Deep Metric Learning

Citation

If you use this method or this code in your research, please cite as:

@InProceedings{Kim_2023_ICCV,
  author = {Kim, Hyungmin and Suh, Sungho and Kim, Daehwan and Jeong, Daun and Cho, Hansang and Kim, Junmo},
  title = {Proxy Anchor-based Unsupervised Learning for Continuous Generalized Category Discovery},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2023},
  pages = {16688-16697}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.