M2A: Motion Aware Attention for Video Action Recognition [ArXiv]

Code is built on top of the TSM module's repo (https://github.com/mit-han-lab/temporal-shift-module).

We use the following code for I3D experiments: https://github.com/IBM/action-recognition-pytorch.

Key Files:

• main.py: Train using DataParralel
• opts.py: command line arguments for training
• ops/sota.py: SOTA temporal modules which we compare against
• ops/temporal_shift.py: M2A, other temporal modules, and inserting the temporal modules into the network
  • CustomMotionAttention is M2A
• scripts/: training script examples