Evaluation on AFLW2000 seems flawed (not comparable to others) about img2pose HOT 1 CLOSED

Hello Felix,

Thank you for your interest in our work!
We are solving a much harder problem than pose estimation, since face detection faces are much harder than faces in pose benchmarks, so we had to create our own labels for training. These labels were created using our own 3D model, thus we tested on AFLW2000-3D using labels created by the ground-truth landmarks and our 3D model.

For approach 1), we compared it with the AFLW2000-3D ground-truth pose labels and got similar results (both single and multi-scale). Because our representation is similar to BIWI, in order to compare to AFLW2000-3D, you need to convert our predicted labels using the function below.

pose_target[:3] = Rotation.from_euler('xyz', pose_para, degrees=False).as_euler('zxy', degrees=True)
pose_pred[:3] = Rotation.from_euler('xyz', convert_to_aflw(pose_pred[:3]), degrees=True).as_euler('zxy', degrees=True)

def convert_to_aflw(rotvec):
    rot_mat_1 = Rotation.from_rotvec(rotvec).as_matrix()
    rot_mat_2 = np.transpose(rot_mat_1)
    angle = Rotation.from_matrix(rot_mat_2).as_euler('xyz', degrees=True)
    return np.array([angle[0], -angle[1], -angle[2]])

For approach 2), although unfeasible since the majority of other methods are not publicly available, a similar conversion would be required.

Hope this helps.

from img2pose.