Skip to content
Conference

Anchor-conditioned bone residual diffusion for 3D human pose estimation

Oct 2026 · International Conference on Image Processing, Object Detection and Tracking · Vol 14506, pp. 145062S - 145062S-6 · 0 citations · 21 references
Engineering

Abstract

Monocular 3D human pose estimation is an inherently ambiguous problem because multiple plausible 3D body configurations may correspond to very similar 2D observations. Existing diffusion-based pose lifting methods improve multihypothesis modeling, but directly denoising the complete 3D pose may increase the learning difficulty and produce unstable local corrections when the 2D input contains detector noise or depth ambiguity. In this paper, we propose Anchor Conditioned Bone Residual Diffusion (ABR-DiffPose), a compact diffusion framework that decomposes 3D pose lifting into stable anchor recovery and structured residual refinement. Specifically, a projection-constrained deterministic anchor first estimates a coarse but stable body pose from the 2D anchor first sequence. The remaining pose error is then transformed into a parent-child bone residual space, where local corrections are represented along the human kinematic tree rather than as independent joint displacements. Finally, an anchor-conditioned dual-stream denoiser performs diffusion refinement in the bone residual space by jointly modeling the stable anchor context and the noisy residual state. Different from simply stacking a deterministic initializer and a Denoising Diffusion Implicit Model (DDIM), the proposed formulation changes the diffusion state itself from full joint coordinates to bounded anchor-relative bone residuals. This design reduces the burden of full-pose diffusion, encourages kinematically coherent refinement, and enables multi-proposal inference for ambiguous 2D-to-3D lifting. Experiments on Human3.6M are conducted to evaluate the effectiveness of the proposed components.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.