Reproducing on ped1&2

[nono-zz]

Hi, thanks for presenting your great work.

I have a problem while reproducing the results on ped2 dataset, the AUC never reached 99% as mentioned in the paper. I got 97% after training, and after fine-tuning, AUC decreased to 95%. Also, it seems like the performance couldn't improve after the first epoch in training.

I strictly followed the preprocessing steps and didn't change the configurations you provided. Do you have any suggestions regarding this issue? And also, have you tested the model on ped1 dataset, I wonder if there're any configurations differ from ped2.

Thank you so much.

[LiUzHiAn]

Hi, thanks for the interest.

I think you have extracted the needed spatial-temporal cubes for training. Could you first load our pre-trained model to see if you can get 99% AUC, in order to check the correctness of the preprocessing steps?

We found the flows are very important for the Ped2 dataset. If your preprocessing is the same as ours, I think you can play with the memory-augmented autoencoder first, trying to get a good reconstruction model. Then, you can go on to stick to the prediction model, and finetuning the hybrid pipeline together. As for the finetuning, we found it improves the final AUC a bit.

We didn't test on Ped1, the foreground objects are much denser, you can tune the object detection thresholds for this dataset. If have other questions, you can drop me an email via liuzhian98@qq.com since I may not be able to reply in time here.

[nono-zz]

Thank you so much for your reply, I'll check that.

[SHAOjav]

Hello This is very excellent work, I am following I want to know how to get this type file (Test002_gt), because I don't find any code about this file. hope to get your reply! Thanks!