inaccurate MANO annotation

[ZhengdiYu]

This is the rendering results of 00005896.jpg of training set. Is this normal?

[lixiny]

I also witnessed a slightly offset on the mask/depth rendering, but it is not as much as your illustration.
Here is my results. ( 00005896.jpg )

img, rendered depth from verts, depth + joints, masks.

[ZhengdiYu]

What kind of rendering method are you using? And what is your MANO model? I use (https://github.com/hassony2/manopth) to get a mesh in camera coordinate system by:

        uv_root, scale = mano[:, 58:60], mano[:, 60:]
        focal, pp = get_focal_pp(K)
        xyz_root = recover_root(uv_root, scale, focal, pp)

        mano_pose = torch.FloatTensor(pose).to(DEVICE)#.view(-1, 3)
        shape = torch.FloatTensor(shape).to(DEVICE)#.view(1, -1)

        assert mano_pose.size() == torch.Size((1, 48)), mano_pose.size()
        assert shape.size() == torch.Size((1, 10)), shape.size()
        output = mano_layer['r'](mano_pose, th_betas=shape) # already root-aligned in manolayer
        joints = output[1][0].cpu().detach().numpy()
        mesh = output[0][0].cpu().detach().numpy()

        trans = xyz_root[0] - joints[9]
        trans = trans[None]
        mesh = mesh + trans
        joints = joints + trans

        # milimeter
        #mesh = (output.vertices[0] * 1000).cpu().detach().numpy() # smplx
        mesh = (mesh * 1000)

[ZhengdiYu]

For most of cases, there are offsets. And of some of them, the offsets are acceptable: