penincillin
about the 3d rotate with 2d rotate
in the data_processor.random_rotate(), the returned joint2d is right:
however the returned joint3d is weird:
and the returned mano output is :
@lvZic How do you visualize the 3D joints and MANO outputs?
i use camera intrinsics to cvt 3D joints to 2D joints (camera to pixel), and use global rotation and position (mano[:3] and mano[-3:]) to cvt MANO outputs to 3D joints.
@lvZic How do you visualize the 3D joints and MANO outputs?
in fact, i found the random rotation strategy is widely used just as your do in data augmentation. But the visual result i saw is not expected. I guees may be the orthographic projection is assumed between 3d and 2d, so the inaccuracy can be removed by the regression output "scale" and trans_x, trans_y"?
@penincillin i now rotate the 3D joints first, then use camera intrinsics to get 2d joints, and use affine transform to rotate image. However, the result seems not just 2d rotation.