RegAD icon indicating copy to clipboard operation
RegAD copied to clipboard

[ECCV2022 Oral] Registration based Few-Shot Anomaly Detection

Results 10 RegAD issues
Sort by recently updated
recently updated
newest added

在该代码的train.py文件中有加载support_set中的pt文件(fixed_fewshot_list = torch.load(f'./support_set/{args.obj}/{args.shot}_{args.inferences}.pt')),但是没有生成pt文件的代码,因此,我想请问您这个pt文件是怎么生成的,代码能公开吗?

在模型训练的时候,如果只有正样本,负样本很少甚至没有,且没有标签数据,该如何修改关于roc_auc_score和test的相关内容呢?希望得到您的回答。

Dear author, thank you very much for your work! However, I am puzzled by the following questions when i trained a model with a custom dataset. 1. If i try...

Can you please share the support set that you used for your MPDD results so that we can have a fair comparison between your method and others? I created a...

1.我发现作者您的代码里train.py文件里有test函数,我想问下这个函数的用处是什么,为什么这里已经测试了,还有一个test.py文件。并且您没有用稳定收敛以后的模型进行保存,而是采用了最好的结果,请问这个结果是不是不具有说服力呢 2.还有就是具体的训练中,查询集是除了所选类别之外的所有类的数据吗还是? 3.您在论文里说只需要训练一个模型,但是在具体的实施过程里,还是每个类训练一个模型,只是训练的方式不一样了,还是说我的理解有问题?

STN的是为了对齐原图,那整个检测过程(包括算support set的均值,协方差,到最终得到anomaly map) 不是应该在STN对其后的特征上面进行吗?代码里面是在逆变换回去后的特征图上进行,那这样STN不是没起作用了吗?望解答,谢谢!

Thank you for your selfless sharing! I see the code below: ` self.transform_x = transforms.Compose([ transforms.Resize([size_tp, size_tp], Image.ANTIALIAS), transforms.CenterCrop([resize, resize]), transforms.ToTensor(), # transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) `...

Is it image level AUC or pixel level? It seems that the paper doesn't tell it explicitly.

May I ask for support_set how is set trained and obtained, and what is the relevant code used

为什么我跑代码的时候,损失total_losses.avg在一开始的时候就很小甚至为负数-0.005116。可以帮我想想原因吗