search

antoine77340 / video_feature_extractor

Easy to use video deep features extractor
Apache License 2.0
310 stars 70 forks source link

Can I extract 3D features per frame? #7

Open Namzakku opened 4 years ago

Namzakku commented 4 years ago

Thanks for the great work! Regarding the question, I did try to change the frame rates within the files and it somehow worked with the 2D features, but for the 3D features, it still show result as per second. Can you show me which value I should change? Thanks!

antoine77340 commented 4 years ago

Did you try to change https://github.com/antoine77340/video_feature_extractor/blob/master/extract.py#L34 and replace 24 by another frame per second value?

Namzakku commented 4 years ago

Thanks for the response! Yes, I tried to change the frame per second value. But the thing is that I found the dimensions of the output seems smaller than what I expected for frame extraction. After reading again your paper and the instructions, I found that

The 3D model is a ResNexT-101 16 frames (https://github.com/kenshohara/3D-ResNets-PyTorch) pretrained on Kinetics. The 3D features are extracted at 1.5 feature per second at the resolution of 112.

So as far as I understand, it might be because of using the ResNexT-101 16 frames model that you have to adjust the shape in here? https://github.com/antoine77340/video_feature_extractor/blob/master/preprocessing.py#L35