Self-Supervised 3D Human Localization using Canonical Pose

초록

Estimating absolute 3D human location from a single RGB image is inherently ambiguous due to inter-subject height variation and pose-induced scale changes. We present a self-supervised canonicalizer that learns, from 2D keypoints only, to (i) map an input pose to a canonical (near-upright) reference and (ii) predict a pose correction scalar that compensates for scale collapse caused by crouching and yaw rotation. Placed before MonoLoco++ as a preprocessing stage, our module stabilizes depth regression by acting as an input normalization mechanism. We benchmark on KITTI and report ALE (Average Localization Error). On KITTI, the All ALE improves from 0.74 m to 0.72 m. Similar gains are observed when training on nuScenes and evaluating on KITTI. The method requires no 3D ground truth and preserves real-time performance. © 2026 IEEE.

키워드