Browse Teach Curate

L3D-Pose: Image-Independent 2D-to-3D Animal Pose Lifting and Avatar Retargeting

A framework that recovers 3D animal pose from a single in-the-wild camera by lifting off-the-shelf 2D keypoints into 3D with a lightweight, image-independent attention-based MLP trained purely on game-engine synthetic data, then retargets the pose onto arbitrary rigged avatars via a lookup table of expanded 'deep-pose' keypoints that add spine and tail landmarks.

KITT Text

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

L3D-Pose (ICASSP 2025) is a framework for recovering 3D animal pose from a single in-the-wild camera and retargeting it onto arbitrary rigged avatars. Because large-scale 3D pose annotations for anim…