WiGesture reproduction — PAD interpolation question

[bobokardashian]

Hi Zijian,

Following your earlier advice, we reproduced CrossFi on WiGesture using process2-split.py + linear interpolation to fill PAD values. Training uses --norm and the full pipeline (full_shot → few_shot).

Current results:
Scenario Ours Paper
In-domain full-shot 72% 98.2%
One-shot cross-domain 39.5% 91.7%
Zero-shot cross-domain 24.4% 64.8%

Our interpolation fills PAD along the time axis per subcarrier, with nearest-neighbor at boundaries. Does this match your approach, or could you share the actual interpolation script? We think this is the main gap.

Minor repo notes: pretrained=True deprecated in newer PyTorch; few_shot.py uses --pretrain_MMD/--test_MMD not --MMD as README states.

Thanks!

[RS2002]

Hi,

Thank you for your efforts in reproducing the work.

Since I left the group that conducted this research two years ago, the data interpolation code has been cleared from the server. I will try to find any backup on my own devices in the coming days.

To determine whether the issue lies in the data processing part, you could first try training ResNet18 (the backbone of our method) in in-domain setting. I recall it achieved over 80% accuracy. You may also try the code in the CV_tasks folder, which was used for some image classification datasets as requested by the paper's reviewers. That code likely does not require any additional data processing.

Best regards,
Zijian