search

dolongbien / HumanBehaviorBKU

Abnormal Human Behaviors Detection/ Road Accident Detection From Surveillance Videos/ Real-World Anomaly Detection in Surveillance Videos/ C3D Feature Extraction
157 stars 57 forks source link

The output score is all 0.The OneDrive link you put is invalid #20

Open daiyuandong opened 4 years ago

daiyuandong commented 4 years ago

Hello, I am running your HumanBehaviorBKU on GitHub. I converted fc-6 into TXT form and put it into the program. The output score is all 0.The OneDrive link you put is invalid, so could you please send me some testing videos (mp4 and fc-6) that you used in the experiment,Or repost a link on GitHub.so that I can check what's wrong with myself.Thank you very much!

dolongbien commented 4 years ago

Hi, Maybe you were feeding the Normal video so that the output score is all zero. Try using abnormal video instead, there are some sample videos under WebApp/media dir. That's for the 0 score problem.

For the link, unfortunately, our work is lost in an unexpected way. I'll try to send you some asap.

daiyuandong commented 4 years ago

Thank you for your help. I used the abnormal video about traffic accident, but the output abnormal scores are all 0. I wonder if it's because I'm using my own video instead of your database so that the video needs to learn first?

dolongbien commented 4 years ago

yea, there is a pre-processing step (resize) before putting into C3D model. Our training videos are fallen under some specific constraints and types of accidents, therefore the model only works well on the test data that has the same distribution (this might be the reason for your 0 score). I think you should try using our sample video first and then apply on yours, finally make sure it's something "like" ours.

daiyuandong commented 4 years ago

I tried to find the UCF crime dataset. Did your dataset apply all the videos of road accidents in the UCF crime dataset? Thank you.Sorry to bother you again.

At 2020-04-20 23:32:37, "Bien Do Long" notifications@github.com wrote:

yea, there is a pre-processing step (resize) before putting into C3D model. Our training videos are fallen under some specific constraints and types of accidents, therefore the model only works well on the test data that has the same distribution (this might be the reason for your 0 score). I think you should try using our sample video first and then apply on yours, finally make sure it's something "like" ours.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

dolongbien commented 4 years ago

No, just a few. But the UCF Crime dataset is all you need, as you are trying to reproduce the feature and score. We take some road accident videos from them.

daiyuandong commented 4 years ago

I'll try. Thank you very much