visualize training data some problems

[Gzzgz]

I converted surreal or h36m data for training. I want to make sure the label of one img is right. so I visualize them.
eg:
for keypoints3d:

this is a res from surreal data. the kpts3d positions misalign to the img.
may be it wrong? What shall I do?

THX

[WenjiaWang0312]

@LazyBusyYang

[LazyBusyYang]

3D keypoints have a different coordinate from 2D image, which means that you cannot overlay locations of a 3d point on img.
If you have the correct camera parameter, here's a possible way.

1. Construct an instance of CameraParameter with K, R, T, H, W.
2. Call camera_parameter.export_to_perspective_cameras() to get a camera in pytorch3d.
3. Project kp3d back to 2d camera spcae. keypoints2d = pytorch3d_perspective_camera.transform_points_screen( points=keypoints3d[..., :3])[..., :2]
4. Visualize kp2d by visualize_kp2d, offering keypoints2d and images at the same time.
   note: It may fail because of wrong cam parameters. If it happens, a simple try with inversed extrinsic might do.

[Gzzgz]

OK.THX