Research map / 3D vision / Human pose estimation

COMPOSE: Hypergraph Cover Optimization for Multi-view 3D Human Pose Estimation

NeurIPS 2026 · 3D vision › Human pose estimation · approach: Monocular 2D-to-3D pose lifting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Recasts multi-view 3D human pose estimation as a weighted exact-cover over a hypergraph of multi-view person hypotheses, solved exactly with ILP or approximately with loopy belief propagation, without any 3D supervision.

Problem

Associating 2D pose detections across sparse calibrated camera views into 3D persons without 3D supervision; pairwise matching followed by synchronization propagates errors under occlusion and noise.

Approach

Treat multi-view person hypotheses (hyperedges grouping at most one detection per view) as atomic associations and select a weighted exact cover over the hypergraph via ILP, with geometric reprojection pruning and a BP relaxation giving soft marginals.

Key results

• On CMU Panoptic, COMPOSE-BP reaches AP25 of 68.88 vs 37.63 for MvPose and 55.13 for SelfPose3d, with COMPOSE-ILP MPJPE 22.78 mm vs MvPose 26.46 mm.
• On Campus, COMPOSE-ILP achieves 97.3% avg PCP, above MvPose (96.3%) and SelfPose3d (87.9%).
• On Shelf, COMPOSE achieves 96.2% avg PCP, above SelfPose3d (95.1%) but below MvPose (96.9%).

Surprising findings

• On Shelf, the proposed method is slightly below MvPose in average PCP (96.2 vs 96.9).

Limitations

• Requires calibrated cameras for triangulation.
• Cannot reconstruct persons visible in only a single view.
• Operates frame-by-frame without temporal information.

Experimental setup

Datasets: CMU Panoptic Shelf Campus

Benchmarks: CMU Panoptic Shelf Campus

Metrics: AP25 AP50 AP100 AP150 Recall@500 MPJPE PCP mAP

Baselines: MvPose 3DPS SelfPose3d DSP VoxelPose + 3DSA Plane Sweep Pose TEMPO Wu et al. ACTOR

Builds on: MvPose 3D Pictorial Structures loopy belief propagation weighted exact cover

Closest papers by meaning

• ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.88)
• Fast and Robust Multi-Person 3D Pose Estimation From Multiple Views (CVPR 2019 · similarity 0.88)
• CanonPose: Self-Supervised Monocular 3D Human Pose Estimation in the Wild (CVPR 2021 · similarity 0.88)
• Multiview-Consistent Semi-Supervised Learning for 3D Human Pose Estimation (CVPR 2020 · similarity 0.88)
• PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space (NeurIPS 2025 · similarity 0.87)
• Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.87)