Dear @cocoon2wong,
Great work on V²-Net — very impressive idea modeling trajectories in the frequency domain! I had a quick question about how frequency components are handled in the implementation.
In the paper, the keypoint estimation sub-network predicts “minimal spectrums” on several “key frequency portions.” However, in the code (e.g., FFTLayer, TrajEncoding), I couldn’t find where:
- Key frequency portions are defined or selected
- The idea of “minimal spectrums” is implemented
- Low-frequency emphasis is explicitly introduced
Are these parts implicitly handled by the architecture, or is there any frequency selection or filtering logic I might have missed?
Thanks a lot for your time, and congrats again on this great contribution!
Best regards,
Dear @cocoon2wong,
Great work on V²-Net — very impressive idea modeling trajectories in the frequency domain! I had a quick question about how frequency components are handled in the implementation.
In the paper, the keypoint estimation sub-network predicts “minimal spectrums” on several “key frequency portions.” However, in the code (e.g., FFTLayer, TrajEncoding), I couldn’t find where:
Are these parts implicitly handled by the architecture, or is there any frequency selection or filtering logic I might have missed?
Thanks a lot for your time, and congrats again on this great contribution!
Best regards,