STARK: A Skeleton-Coupled Transition Autoregressive Kernel for Multi-Hypothesis Monocular 3D Human Pose Estimation

Video-based motion analysis has made single-camera 3D pose a practical biosignal for clinical gait assessment. However, it is inherently ambiguous, as multiple 3D configurations can produce nearly identical 2D poses. Recent conditional generative methods model this one-to-many mapping by transforming Gaussian noise into plausible 3D poses conditioned on the observed 2D pose. Among them, diffusion models require many denoising steps, whereas flow matching performs the transport in only a few steps. However, the standard flow-matching objective predicts the conditional mean of the transport displacement. Under coarse integration, this mean-valued update can underrepresent competing depth modes and produce insufficiently dispersed hypothesis sets. In this paper, we propose STARK for multi-hypothesis monocular 3D human pose estimation. STARK leverages Transition Matching to sample the conditional transport displacement rather than its mean, allowing different samples to capture different plausible depth modes. To maintain whole-body consistency, STARK samples joint transitions autoregressively along the kinematic tree, conditioning each joint on previously sampled joints and the current pose. Experiments show that STARK achieves state-of-the-art coverage among comparable probabilistic methods in the multi hypothesis setting on Human3.6M. STARK also remains competitive in single prediction and cross-dataset settings, generalizing to MPI-INF-3DHP without fine-tuning.