3D coordinate key points of hand 21

[cgz6498]

Where is the 3D coordinate output of key points of hand 21 in demo?

[mks0601]

You can get 3D keypoint coordinates from mesh

InterWild/demo/demo.py

Line 160 in 7c01e4a

save_obj(mesh, mano.face[h], osp.join(mesh_save_path, file_name + '_' + h + '.obj'))

using the regressor

InterWild/main/model.py

Line 47 in 7c01e4a

joint_cam = torch.bmm(torch.from_numpy(mano.sh_joint_regressor).cuda()[None,:,:].repeat(batch_size,1,1), mesh_cam)

[cgz6498]

Thank you, but what I want to get is the 3D keypoint coordinates of the hand relative to the input image. Using the method you provided, the generated 3D coordinates are all less than 1. It does not match the 2D coordinates of joint_img in the demo, as shown in the figure.

At the same time, please ask the joint_img of the following figure to obtain the 3D coordinates, and then obtain the 2D coordinates relative to the input image through coordinate conversion. Can you also get the 3D coordinates relative to the picture through the coordinate transformation and show them?

[mks0601]

I'm not clear. What do you mean by 3D coordinates relative to the input image? 3D coordinates in camera-centered coordinates in meter scale? if so, what I answered is the correct one.

[cgz6498]

My aim is to visualize the 3D coordinates of gestures, with depth (z) marked in different colors. Visualization of 2D coordinates similar to the figure below.