Awesome

Rotation Regularization Without Rotation

The Pytorch implementation for the ECCV2022 paper of "Rotation Regularization Without Rotation" by Takumi Kobayashi.

Citation

If you find our project useful in your research, please cite it as follows:

@inproceedings{kobayashi2022eccv,
  title={Rotation Regularization Without Rotation},
  author={Takumi Kobayashi},
  booktitle={Proceedings of European Conference on Computer Vision (ECCV)},
  year={2022}
}

Usage

Training

For example, the ResNet-10 is trained from scratch with our regularization on ImageNet-LT dataset by

CUDA_VISIBLE_DEVICES=0,1 python main_train.py --dataset imagenetlt  --data ./datasets/imagenetlt --arch ResNet10Feature  --epochs 180 --out-dir ./results/imagenetlt/ResNet10Feature/train_1st_stage --workers 12 

CUDA_VISIBLE_DEVICES=0,1 python main_train.py --dataset imagenetlt  --data ./datasets/imagenetlt --arch ResNet10Feature  --epochs 30 --out-dir ./results/imagenetlt/ResNet10Feature/train_2nd_stage --workers 12 --first-model-file ./results/imagenetlt/ResNet10Feature/train_1st_stage/model_best.pth.tar

Note that the ImageNet-LT dataset must be downloaded at ./datasets/imagenetlt/ before the training and here we follow the imbalance-aware 2-stage training procedure presented in [1].

You can also apply our regularization to the framework of logit adjustment [2] by

CUDA_VISIBLE_DEVICES=0,1 python main_train.py --dataset imagenetlt  --data ./datasets/imagenetlt --arch ResNet10Feature --logit-adjust  --epochs 180 --out-dir ./results/imagenetlt/ResNet10Feature/logit_adjust --workers 12 

Results

ImageNet

References

[1] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng and Yannis Kalantidis. "DECOUPLING REPRESENTATION AND CLASSIFIER FOR LONG-TAILED RECOGNITION." In ICLR, 2020.

[2] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit and Sanjiv Kumar. "LONG-TAIL LEARNING VIA LOGIT ADJUSTMENT." In ICLR, 2021.

Contact

takumi.kobayashi (At) aist.go.jp

Acknowledgement

The class-wise sampler utils/ClassAwareSampler.py is from the Classifier-Balancing.