L3D-Pose: Lifting Pose for 3D Avatars from a Single Camera in the Wild

3D pose estimation helps analyze animal movements but is limited by the difficulty of creating datasets. We address this with a hybrid approach using rigged avatars to generate synthetic 3D data. Our lightweight attention-based MLP converts 2D poses to 3D and retargets wild movements onto 3D avatars.

Go to the Official Project Page