Hi, thank you for sharing your PoC code. I'm interested in Anomaly Detection, so found your paper and code. Based on your PoC code and some own refine, after trying to reproduce the results of the UCSD Peds1 dataset, the original paper results could not be reproduced at all (EER in FL/PL is about 50 %).

I have some questions.

1. How was the validation data used to determine the timing for saving model parameters splitted? Did you split Train into train/val or Test directly into val?

2. In the "Stopping criteria" section of the original paper, there is a description that the Inpainter model parameters are saved when the reconstruction loss is in the minimum point. However, the loss function of AVID in (3) equation dose not include the reconstruction loss. Are these not contradiction?

3. What does the square operation in the loss function of AVID in (3) equation mean? I think this operation is meaningless since the log(x^2) is equivalent to 2*log(x) and acts as a constant coefficient.

4. How was the pixel value threshold determined in the original paper results? And what was the specific threshold? (the value defined in (4) equation as notation alpha)

Thanks

For the snippet,
transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5)),
it should be changed for the images with gray scale like below.

transforms.Normalize((0.5,), (0.5,)),

Please change the citation to the ACCV'18 published paper on the springer. Also, update the README file about the final status of the paper.

I've trained the model for 40 epochs. The generated image shows some weird artifacts as shown below. What could be wrong? I did not modify the model.

Change the title to the paper title: "AVID: Adversarial Visual Irregularity Detection"