johnnyhopp / padim-efficientnetv2 Goto Github PK

Hello, I would like to ask a question. Now I have found that the Gaussian distribution at certain points exhibits a double Gaussian distribution, and there may be deviations when calculating the mean and covariance. I would like to distinguish the channels that conform to the double Gaussian distribution, and then calculate the minimum Mahalanobis distance separately from the two distributions. Do you have any ideas about this?

After calculating mean and cov for each patch(x, y) in training set, I feed an image through pretrained model, take out feature maps of each layer, concatenate them then calculate distance to each patch distribution. Basically, I will have the score_map, then up-sample it and normalize it. After that, how can I tell if input data is defected or not? Assume that I already have a optimal threshold. I compare this score_map to the threshold. If any pixel of score_map is bigger than the threshold value, input data will be predicted as defect. In the other hand, if all pixels of score_map are smaller than threshold value, it will be predicted as normal. Am I correct? If not, please let me know. Many thanks!

Hey Johnny, thanks for sharing your project!

I was wondering if you are able to provide at some point in time, a predict.py script.

So we can infer anomalies on single images, in a real life scenario, after the training stage is

completed.

On this video, you can see an implementation of Padim https://github.com/OpenAOI/anodet.

https://www.youtube.com/watch?v=W4-ZtpJtE5c&ab_channel=tektronix475

I would like to do the same tests, with your PaDiM-EfficientNetV2

Thank you.