WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation, CVPR 2023 [Paper]
If you find the code useful, please consider citing our paper using the following BibTeX entry.
@InProceedings{Jeong_2023_CVPR,
author = {Jeong, Jongheon and Zou, Yang and Kim, Taewan and Zhang, Dongqing and Ravichandran, Avinash and Dabeer, Onkar},
title = {WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {19606-19616}
}
@misc{cao2023segment,
title={Segment Any Anomaly without Training via Hybrid Prompt Regularization},
author={Yunkang Cao and Xiaohao Xu and Chen Sun and Yuqi Cheng and Zongwei Du and Liang Gao and Weiming Shen},
year={2023},
eprint={2305.10724},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
- Python 3.7, PyTorch 1.10, and more in install.sh
sh install.sh
- Follow instructions in https://www.mvtec.com/company/research/datasets/mvtec-ad/
- Follow instructions in https://paperswithcode.com/dataset/visa
python run_winclip.py
MVTec-AD | Reported | Re-implementation | ||||||
---|---|---|---|---|---|---|---|---|
i-auroc | p-auroc | i-max-f1 | p-max-f1 | i-auroc | p-auroc | i-max-f1 | p-max-f1 | |
carpet | 100.00 | 95.40 | 99.40 | 49.70 | 77.41 | 88.96 | 88.44 | 29.31 |
grid | 98.80 | 82.20 | 98.20 | 18.60 | 48.87 | 75.08 | 85.71 | 8.40 |
leather | 100.00 | 96.70 | 100.00 | 39.70 | 97.35 | 97.35 | 95.70 | 29.60 |
tile | 100.00 | 77.60 | 99.40 | 32.60 | 79.87 | 75.87 | 85.25 | 29.30 |
wood | 99.40 | 93.40 | 98.30 | 51.50 | 94.74 | 93.03 | 92.68 | 44.65 |
bottle | 99.20 | 89.50 | 97.60 | 58.10 | 98.65 | 89.58 | 96.77 | 49.36 |
cable | 86.50 | 77.00 | 84.50 | 19.70 | 53.30 | 56.23 | 76.03 | 10.22 |
capsule | 72.90 | 86.90 | 91.40 | 21.70 | 62.03 | 88.56 | 90.46 | 9.95 |
hazelnut | 93.90 | 94.30 | 89.70 | 37.60 | 71.29 | 94.34 | 80.00 | 33.63 |
metal_nut | 97.10 | 61.00 | 96.30 | 32.40 | 37.59 | 42.67 | 89.42 | 21.67 |
pill | 79.10 | 80.00 | 91.60 | 17.60 | 73.10 | 74.67 | 91.56 | 11.98 |
screw | 83.30 | 89.60 | 87.40 | 13.50 | 64.87 | 90.09 | 85.61 | 9.09 |
toothbrush | 87.50 | 86.90 | 87.90 | 17.10 | 41.94 | 84.02 | 84.51 | 9.26 |
transistor | 88.00 | 74.70 | 79.50 | 30.50 | 62.25 | 67.46 | 60.87 | 15.95 |
zipper | 91.50 | 91.60 | 92.90 | 34.40 | 89.31 | 92.08 | 90.42 | 31.48 |
Average | 91.81 | 85.12 | 92.94 | 31.65 | 70.17 | 80.67 | 86.23 | 22.92 |
VisA | Reported | Re-implementation | ||||||
---|---|---|---|---|---|---|---|---|
i-auroc | p-auroc | i-max-f1 | p-max-f1 | i-auroc | p-auroc | i-max-f1 | p-max-f1 | |
candle | 95.40 | 88.90 | 89.40 | 22.50 | 79.03 | 86.24 | 72.36 | 6.32 |
capsules | 85.00 | 81.60 | 83.90 | 9.20 | 53.58 | 62.00 | 77.22 | 1.36 |
cashew | 92.10 | 84.70 | 88.40 | 13.20 | 70.66 | 79.54 | 80.99 | 6.94 |
chewinggum | 96.50 | 93.30 | 94.80 | 41.10 | 84.94 | 97.01 | 83.76 | 36.17 |
fryum | 80.30 | 88.50 | 82.70 | 22.10 | 52.60 | 86.73 | 80.33 | 15.17 |
macaroni1 | 76.20 | 70.90 | 74.20 | 7.00 | 49.98 | 34.37 | 66.67 | 0.07 |
macaroni2 | 63.70 | 59.30 | 69.80 | 1.00 | 49.56 | 31.49 | 66.67 | 0.06 |
pcb1 | 73.60 | 61.20 | 71.00 | 2.40 | 55.99 | 44.04 | 68.97 | 0.97 |
pcb2 | 51.20 | 71.60 | 67.10 | 4.70 | 61.58 | 64.47 | 69.26 | 0.70 |
pcb3 | 73.40 | 85.30 | 71.00 | 10.30 | 51.42 | 68.71 | 66.45 | 1.06 |
pcb4 | 79.60 | 94.40 | 74.90 | 32.00 | 78.94 | 91.86 | 74.56 | 22.75 |
pipe_fryum | 69.70 | 75.40 | 80.70 | 12.30 | 82.80 | 93.65 | 83.48 | 22.45 |
Average | 78.06 | 79.59 | 78.99 | 14.82 | 64.26 | 70.01 | 74.23 | 9.50 |
This project borrows some code from OpenCLip and CDO, thanks for their admiring contributions~!