search

sungwool / CFA_for_anomaly_localization

103 stars 22 forks source link

Inference Speed and CFA++? #1

Closed mjack3 closed 2 years ago

mjack3 commented 2 years ago

Hello and congrat for this work.

I wonder if you studied the inference time. Paper doesn't say anything about that.

mjack3 commented 2 years ago

I would like to know what is cfa++. I don't catch the idea of the paper.

sungwool commented 2 years ago

Thank you for your interest.

A. Table 6 presents the throughput of the proposed method. Throughput refers to the number of images that can be referenced in one second.

B. In some classes in MVTec AD, noise in the background may interfere with detection. Therefore, I used the results obtained by simultaneously using both cropped and uncropped images. Among the performances obtained from the two images, 'CFA++' is the only the best performance. In general, anomaly detection performs well on cropped images and anomaly localization on uncropped images.

Thanks.

mjack3 commented 2 years ago

Thanks for your answer.

Please, I'd like to do a last question.

Why do you use Image.Antialias? As far as I know, this is enabled with the antialias flag of Resize (by default) if image is a PIL. I think that this is your case.

Thanks

sungwool commented 2 years ago

Hi,

There is no particular reason for it. I just wanted to make sure that 'anti-aliasing' is activated by specifying the flag :)

Thanks!

William9Baker commented 1 year ago

Thank you for your interest.

A. Table 6 presents the throughput of the proposed method. Throughput refers to the number of images that can be referenced in one second.

B. In some classes in MVTec AD, noise in the background may interfere with detection. Therefore, I used the results obtained by simultaneously using both cropped and uncropped images. Among the performances obtained from the two images, 'CFA++' is the only the best performance. In general, anomaly detection performs well on cropped images and anomaly localization on uncropped images.

Thanks.

But I can't find the implementation of CFA++ in the code.

mjack3 commented 1 year ago

CFA does not exist. It is just a cherry-picking results between cropped and entire images.