sharonn18

/

tgcn-wlasl

like 0

Hi,
I’m using your TGCN-WLASL model for sign language recognition and need to match the exact keypoint format used during training. The model expects input of shape (batch, 55, num_samples*2), but the order of the 55 keypoints is not documented. I’d like to clarify :

1. Order of keypoint groups: Is it body (13) + left_hand (21) + right_hand (21), or another order ?
2. Body subset: Which body keypoints are used?
3. Hand keypoints: Are hand keypoints in OpenPose hand format (21 points per hand)? If so, what is the exact order (e.g. wrist, thumb CMC→tip, index MCP→tip, etc.)?
4. Coordinate system: Are coordinates in pixel space (x, y in image dimensions) or normalized (e.g. 0–1)? Is there centering or scaling applied?
5. Temporal format: For each keypoint, is the feature vector ordered as [x_t1, y_t1, x_t2, y_t2, ...] or [x_t1..x_t50, y_t1..y_t50]?

If you have any preprocessing script that shows how WLASL keypoints are converted to the model input, that would be very helpful.
Thank you for your work and for any guidance you can provide.
Best regards,
Aiana