WinCLIP: 零/少样本异常分类和分割

<p align="center"><img src="https://yellow-cdn.veclightyear.com/ab5030c0/821cdc28-7a7a-46fd-87aa-60575c23dd61.jpg" alt="outline" width="90%"></p>

非官方实现:

WinCLIP: 零/少样本异常分类和分割, CVPR 2023 [论文]

引用

如果您发现这个代码很有用,请考虑使用以下BibTeX条目引用我们的论文。

@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}
}

相关研究

@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}
}

相关仓库

• Segment Any Anomaly

前提条件

• Python 3.7, PyTorch 1.10, 和 install.sh中更多的内容

安装python依赖

sh install.sh

下载MVTec-AD数据集

• 按照 https://www.mvtec.com/company/research/datasets/mvtec-ad/ 中的指示进行

下载Visa数据集

• 按照 https://paperswithcode.com/dataset/visa 中的指示进行

运行 run_winclip.py 以复现实现结果

python run_winclip.py

结果

MVTec-AD

MVTec-AD	报告				重新实现
	i-auroc	p-auroc	i-max-f1	p-max-f1	i-auroc	p-auroc	i-max-f1	p-max-f1
地毯	100.00	95.40	99.40	49.70	77.41	88.96	88.44	29.31
网格	98.80	82.20	98.20	18.60	48.87	75.08	85.71	8.40
皮革	100.00	96.70	100.00	39.70	97.35	97.35	95.70	29.60
瓷砖	100.00	77.60	99.40	32.60	79.87	75.87	85.25	29.30
木材	99.40	93.40	98.30	51.50	94.74	93.03	92.68	44.65
瓶	99.20	89.50	97.60	58.10	98.65	89.58	96.77	49.36
电缆	86.50	77.00	84.50	19.70	53.30	56.23	76.03	10.22
胶囊	72.90	86.90	91.40	21.70	62.03	88.56	90.46	9.95
榛子	93.90	94.30	89.70	37.60	71.29	94.34	80.00	33.63
金属螺母	97.10	61.00	96.30	32.40	37.59	42.67	89.42	21.67
药片	79.10	80.00	91.60	17.60	73.10	74.67	91.56	11.98
螺丝	83.30	89.60	87.40	13.50	64.87	90.09	85.61	9.09
牙刷	87.50	86.90	87.90	17.10	41.94	84.02	84.51	9.26
晶体管	88.00	74.70	79.50	30.50	62.25	67.46	60.87	15.95
拉链	91.50	91.60	92.90	34.40	89.31	92.08	90.42	31.48
平均数	91.81	85.12	92.94	31.65	70.17	80.67	86.23	22.92

VisA

VisA	报告				重新实现
	i-auroc	p-auroc	i-max-f1	p-max-f1	i-auroc	p-auroc	i-max-f1	p-max-f1
蜡烛	95.40	88.90	89.40	22.50	79.03	86.24	72.36	6.32
胶囊	85.00	81.60	83.90	9.20	53.58	62.00	77.22	1.36
腰果	92.10	84.70	88.40	13.20	70.66	79.54	80.99	6.94
口香糖	96.50	93.30	94.80	41.10	84.94	97.01	83.76	36.17
炸薯条	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
平均	78.06	79.59	78.99	14.82	64.26	70.01	74.23	9.50