Transfer learning on action recognition


I’m testing transfer learning based on a ResNext3D model using a custom dataset.
I have two questions:

  1. Is the same frame rate required in my dataset? THe UCF101 human action videos uses 25 FPS (and 320×240 resolution)[1]. I used 10 frames per second until now to have a smaller total dataset size.

  2. Is the ResNext3D model the best to use as a base for transfer learning in action recognition? Is there a model with better performance released?

Thanks a lot for any answers, links or tips! :slight_smile:

[1] Original UCF101 Paper: