Access estimated transformation of mesh

[mjvanderboon]

Hi, thanks for the great work and repository.

I am looking to access the estimated homogeneous transform that aligns the mesh from an initial frame to a frame that is aligned with the camera position of an input image. Or conversely the homogeneous transform that transform the camera pose to align the mesh with the image.

I suspect that the rotational information I need is available as an angle axis representation in the first pose entry in the code_dict: code_dict['pose'][..., 0:3], could you confirm this is correct? I also suspect the translation vector I desire is available as code_dict['cam'], but I am unsure how to transform this [scale, t_x, t_y] vector into a translation vector in R3.

Could you provide any hints or pseudocode on how to generate a homogeneous transformation matrix from the available pose estimation?

[FeiiYin]

Can the tform in dataset works?

[mjvanderboon]

I looked at the code in /decalib/datasets/datasets.py and I would be able to write a similar functionality to extract the estimated transformation for a general input image using a face detector. However, as far as I understand code_dict['pose'] should also encode the pose of the face together with code_dict['cam']. Do you know if this is correct? Or do these values encode other parameters?

[mjvanderboon]

A bit more detail is available in the issue linked above. It explains how to use code_dict['pose'] and code_dict['cam'] to estimate full homogeneous transformation matrices.