GithubHelp home page GithubHelp logo

xrli-u / musc Goto Github PK

View Code? Open in Web Editor NEW
223.0 223.0 12.0 8.41 MB

This is an official PyTorch implementation for "MuSc : Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images" (MuSc ICLR2024).

License: MIT License

Python 94.21% Shell 5.79%

musc's People

Contributors

xrli-u avatar

Stargazers

 avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar  avatar

Watchers

 avatar  avatar  avatar

musc's Issues

how to use dino or dinov2?

FileNotFoundError: [Errno 2]No such file or directory: '/home/robot/.cache/torch/hub/facebookresearch dinov2 main/hubconf.py'

musc模型进一步细节

利用用未标记测试图像,提出了一种基于未标记图像相互评分进行分类与分割,这确实是一种新机制。

我想请问:这是属于:无监督及自监督的思路吧?(通常是利用:正常合格的产品,即正样本去训练。)

从论文中,我的感悟,你这个思路,好像是类似:基于合成的模型?提取VIT特征,然后分区域打上标签?然后把合成前后的图像,互相打分?

请问:引用Dino V2作为骨干,是否效果最优呢?

另外 :你论文中提到的RsCIN 有3个局限性:

1:对其骨干网中不包含图像级特征的一些单类方法,需要另一个骨干网来提取图像级特征。

2:对于分类结果非常高的方法,RsCIN 对图像级特征表示的改进(潜力空间)有限?

3:你们发现,你们的方法和 WinCLIP 都受到方向或比例不一致的影响。原因是使用固定的预训练视觉Transformer 作为特征提取器,在预训练过程中,Transformer 只对方向和比例进行了轻微的数据增强。因此,我们都不能有效地解决方向或比例不一致的问题。APRIL-GAN 的 AC 分数也有所下降,但 AS 分数有所提高。猜测原因是 APRIL-GAN 使用额外的训练集来优化 AS。

我很有兴趣你们团队是如何持续改进,解决以上问题的。当然,这样就离实际工业化落地更进一步了。

btad和mvtec_anomaly_detection无法可视化

visa可以直接可视化结果。可是剩余的两个数据集在结果文件夹只空的文件夹。请问是参数设置出现什么问题了吗?以下是我的参数设置。
parser = argparse.ArgumentParser(description='MuSc') parser.add_argument('--config', type=str, default='./configs/musc.yaml', help='config file path') parser.add_argument('--data_path', type=str, default="./data/mvtec_anomaly_detection/", help='dataset path') parser.add_argument('--dataset_name', type=str, default="mvtec_ad", help='dataset name') parser.add_argument('--class_name', type=str, default="ALL", help='category') parser.add_argument('--device', type=int, default=0, help='gpu id') parser.add_argument('--output_dir', type=str, default="./output/", help='save results path') parser.add_argument('--vis', type=str, default="true", help='visualization') parser.add_argument('--vis_type', type=str, default="single_norm", help='normalization type in visualization') parser.add_argument('--save_excel', type=str, default=None, help='save excel') parser.add_argument('--r_list', type=int, nargs="+", default=None, help='aggregation degrees of LNAMD') parser.add_argument('--feature_layers', type=int, nargs="+", default=None, help='feature layers') parser.add_argument('--backbone_name', type=str, default=None, help='backbone') parser.add_argument('--pretrained', type=str, default=None, help='pretrained datasets') parser.add_argument('--img_resize', type=int, default=224, help='image size') parser.add_argument('--batch_size', type=int, default=32, help='batch size') parser.add_argument('--divide_num', type=int, default=None, help='the number of divided subsets')

作者大大,请问你可以帮忙我们做gui界面吗?

我想装在win 系统上面,pyhone 语言,我可以有偿付费给你,目的是方便测试我自己的数据集,以便输入,输出,调参等,谢谢。考虑到软件由你们开发,所以,一些细微的接口等,你们轻车熟路。 我不要求美观,实用就行。

然后,我们试跑后,会把实际结果反馈与你,以便你们进一步优化改进,谢谢。我们真的也很看好这个模型的潜力,然后,我们会用实际工业中的场景去验证,我想这样的话, 更加会让你们架构接近完美。

feature extractors

Very good work! Do you support other feature extractors, or can you only choose from these three?

about figure 2 visualization

Very good work!!! How is the heat map in the bottom right corner of Figure 2 visualized in the article? Can you explain or provide relevant code? Thank you!!

Results question

when I use your default setting in configs/musc.yaml (e.g. vit-l-14-336, batch_size: 4, feature_layers: [5, 11, 17, 23], r_list: [1, 3, 5]), I only get the results: 96.8 96.6 98.8 97.1 62.2 62.3 93.5 (auroc_sp f1_sp ap_sp auroc_px f1_px ap_px aupro)

A little question

Why is the test image scaled to a resolution size of 518*518, is there any reason or prior knowledge?

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.