Inclusive Hand Pose Estimation via Topology-Aware Distillation and Progressive Kinematic Reasoning

Abstract

Hand pose estimation aims to recover the spatial configuration of hand joints from visual observations and serves as a fundamental perception module for downstream systems that rely on complete and topologically consistent hand representations. Most existing methods are developed under the assumption of fully articulated hand structures and tend to produce unstable or implausible predictions when parts of the kinematic structure are persistently absent, resulting in degraded performance or even system failure despite the availability of partial observations. To systematically investigate this challenge, we construct IncNYU, a physiologically constrained benchmark that simulates structural incompleteness through morphological modification of complete hand models and retains reference locations for both intact and synthetically removed regions. Based on this benchmark, we propose InsHaPE, an Inclusive Hand Pose Estimation framework designed to remain robust under structural incompleteness while preserving performance on standard, fully articulated hands. InsHaPE is built upon a hierarchical kinematic modeling paradigm and introduces two complementary components. First, a topology-aware distillation strategy transfers joint-level and relational priors from fully articulated hand representations, providing anatomically consistent initialization under missing structures. Second, a progressive kinematic reasoning module performs structured inference along the kinematic chain, enabling complete and anatomically consistent pose estimation under severe structural variations. Extensive experiments on NYU/IncNYU, together with external validation on MSRA, demonstrate that InsHaPE improves robustness under structural incompleteness while maintaining competitive performance on complete hands. Additional system-level evaluations further show that InsHaPE improves robustness for downstream applications that require fixed-topology hand pose representations. Our code is available at https://github.com/tomhao555/InsHaPE.