Official PyTorch Repository of "Difficulty-Aware Simulator for Open Set Recognition" (ECCV 2022 Paper)
- python 3.6
- torch 1.2
- torchvision 0.4
- CUDA 10.1
- scikit-learn 0.22
Split information for all datasets can be found in split.py
.
splits_F1
and splits_AUROC
are split information for each benchmark with F1 score and AUROC.
When you run the code, datasets except tiny-ImageNet will be automatically downloaded.
python osr.py --dataset 'cifar10'
To run the code, execute osr.py
.
Then, the results will be saved under the "logs" directory.
sh osr.sh
For simplicity, we provide the training scripts for running all datasets. You can execute the shell file by the above command.
Since [Open-Set Recognition: a Good Closed-Set Classifier is All You Need?] elaborated additional techniques with searched hyperparameters can boost OSR performances, we here simply conduct and compare the performances of DIAS and ARPL cs.
Tiny-ImageNet | 800 epochs training |
---|---|
ARPL cs | 71.9 |
DIAS (Ours) | 75.6 |
If you find this repository useful, please use the following entry for citation.
@inproceedings{moon2022difficulty,
title={Difficulty-Aware Simulator for Open Set Recognition},
author={Moon, WonJun and Park, Junho and Seong, Hyun Seok and Cho, Cheol-Ho and Heo, Jae-Pil},
booktitle={European Conference on Computer Vision},
year={2022},
organization={Springer}
}
If there are any questions, feel free to contact with the authors: WonJun Moon ([email protected]), JunHo Park ([email protected]), Hyun Seok Seong ([email protected]), Cheol-Ho Cho ([email protected]).
This repository is built based on ARPL repository. Thanks for the great work.