OPear

/

videomae-large-finetuned-UCF-Crime

like 13

Follow

OPear Vision 11

Why can't you finetune your ucf crime dataset on VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. It would give better results.

Thanks for your feedback, we actually didn't realize this model existed!

Hey @Shinoy17 ! Would you like to share the approach you would take I think you have great context on this topic and it would be very helpful

Hi, I have tried OPEAR, but it is giving wrong results in the normal videos; it says "Abuse" to them. If you want I can send the link of the normal videos that I crosscheck too.

Hey @p-a-b , I just sent an email with the video links along with the full database I used for cross-checking. Thanks a lot for your insights.

I have the same issue with wrong labels. Seems to be defaulting to abuse or burglary in perfectly normal video. Can you share what the diagnosis from Retro0Kivi's problem was?

Hey @ddavidkov we were waiting to check the false positives from Retr0Kivi to understand the issues but we did not receive the videos.

To be able to debug and diagnose the issue.

Would you mind uploading them to a cloud, including tag and confidence?

I am sorry that I could not send you the videos because my computer went down and I was trying to fix it and also some issues with my master's but I handled all of them. I could not made a file dedicated to wrong result videos but you can find them here "https://www.kaggle.com/datasets/minhajuddinmeraj/anomalydetectiondatasetucf?select=Testing_Normal_Videos_Anomaly". These normal videos and also other normal videos from this dataset, are labeled as abuse instead of normal.

I am also having trouble also with understanding the normal videos problem. Every normal video from the dataset is mislabeled. Why does this happen?