How to make face model straight front as in the gif ? about deep3dfacerecon_pytorch

[YuDeng]

Hi, the rotation stores the three Euler angles (in rad) w.r.t to three rotation axis (x,y,z axis). The translation stores the xyz shift (in dm) w.r.t the origin in the world coordinate space. For more details, you can check https://github.com/sicxu/Deep3DFaceRecon_pytorch/blob/master/models/bfm.py

[YuDeng]

Hi, to generate obj file with frontal face, you need to change the predicted pose parameters to zeros.
You can modify the angles and translation here:

Deep3DFaceRecon_pytorch/models/bfm.py

Line 263 in 73d4911

angles = coeffs[:, 224: 227]

[Michaelwhite34]

Hi, to generate obj file with frontal face, you need to change the predicted pose parameters to zeros.
You can modify the angles and translation here:

Deep3DFaceRecon_pytorch/models/bfm.py

Line 263 in 73d4911

angles = coeffs[:, 224: 227]

Thanks for you reply! Do I just need to use "angles = coeffs[0, 0: 0]" ?

[Michaelwhite34]

Also,assuming the face is front in the 3d scene,how to output the camera data that is used for the overlay comparison (i.e the camera position rotation and focal length) ? （P.S. sorry for the stupid questions,I am not really familar with all this）

[YuDeng]

You just simply set angles = coeffs[:, 224: 227] *0 would be fine.

In our work we set the camera position at [0,0,10] with a fixed focal length of 1015. And the camera is looking at [0,0,0] which is the origin of the world coordinate. The predicted angles and translation in "coeff" are rotation and shift for the 3d face w.r.t the world coordinate. With the above information, you should be able to render a face image aligned with the input face.

[Michaelwhite34]

Thank you ! btw by saying the camera information I mean the "transformed" information (the camera look at the face so it will rotate as the face rotate to front view)..Can I output that information to cmd window ? (I know it can be computed from the face rotation and translation but direct output would be convenient)

[YuDeng]

In our case we do not record the camera extrinsic parameters but only face rotations and translations. So I think you may need to compute it by yourself.

[Michaelwhite34]

That's alright,I can do this by camera tracking in blender.Then can I output the translation and rotation of face to cmd window (in a manner that (x,y,z) for translation and another(x,y,z) for rotation)?

[YuDeng]

The default output already contains these information. If you check the test section, the output .mat file contains the "coeffs". If you just want to print them in the cmd window, you can modify the function here:

Deep3DFaceRecon_pytorch/test.py

Line 68 in 73d4911

model.save_coeff(os.path.join(visualizer.img_dir, name.split(os.path.sep)[-1], 'epoch_%s_%06d'%(opt.epoch, 0),img_name+'.mat')) # save predicted coefficients

I'm not sure if this is what you expect.

[Michaelwhite34]

What’s the exact meaning of array in rotation and translation ? Do they mean sin value of angle and meter in distance respectively ?

[Michaelwhite34]

The default output already contains these information. If you check the test section, the output .mat file contains the "coeffs". If you just want to print them in the cmd window, you can modify the function here:

Deep3DFaceRecon_pytorch/test.py

Line 68 in 73d4911

model.save_coeff(os.path.join(visualizer.img_dir, name.split(os.path.sep)[-1], 'epoch_%s_%06d'%(opt.epoch, 0),img_name+'.mat')) # save predicted coefficients

I'm not sure if this is what you expect.

Could you tell me the exact meaning of the array stored in rotation and translation?

[Michaelwhite34]

Hi, the rotation stores the three Euler angles (in rad) w.r.t to three rotation axis (x,y,z axis). The translation stores the xyz shift (in dm) w.r.t the origin in the world coordinate space. For more details, you can check https://github.com/sicxu/Deep3DFaceRecon_pytorch/blob/master/models/bfm.py

I did an experiment about translation and rotation.I have 3 meshes,one disabled rotation(1),one disable translation(2) and one disabled both(3).I try to match them with the given rotation and translation parameter.
(left is rotation right is translation)
(mesh 1 and mesh 3 coincide with output translation)

However mesh 2 and mesh 3 won't even be close with the output rotation(transfer from rad to degree which is *180/pi).I try to align them manually and get the rotation (-5.2,7.63,42.8)(in degree) (not perfect but very close)...Do you know what's going on here ?

[Michaelwhite34]

hi,how can I export the original bounding box without resizing to 224*224 ? I want to use that for texture projection.