Research map / 3D vision / Human pose estimation

ProPose: Probabilistic 3D Human Pose Estimation with Instance-Level Distribution and Normalizing Flow

AAAI 2025 · 3D vision › Human pose estimation · approach: Transformer-based multi-view 3D pose · 1 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.

Proposes ProPose, a probabilistic 3D human pose lifting framework that maps a pose-instance-level Gaussian (PILGD) to a 2D pose-adaptive Gaussian (PAGD) via a conditional normalizing flow, with self-representation autoencoders and a discriminator.

Problem

Lifting 2D poses to 3D is one-to-many; prior generative multi-hypothesis methods sample from a pose-independent standard Gaussian even though depth distributions vary by pose, leading to inefficient distribution learning and implausible 3D poses.

Approach

Replace the standard Gaussian prior with a pose-instance-level Gaussian learned by a conditional autoencoder, and use a 2D-pose-adaptive Gaussian as the normalizing-flow base distribution, so 3D estimation becomes a mapping between two Gaussians.

Key results

• In the multi-hypothesis setting on H3.6M, ProPose with S=200 reaches 36.4 mm under protocol 1 and 28.8 mm under protocol 2, comparable to prior state-of-the-art probabilistic methods.
• In the single-hypothesis setting on H3.6M protocol 1, ProPose reaches 51.9 mm MPJPE, comparable to deterministic methods.
• On MPI-INF-3DHP cross-dataset evaluation (trained on H3.6M), ProPose reports 52.1 AUC and 83.9 PCK overall in the SH setting, above the compared probabilistic and deterministic methods listed.

Measured effects reported in the paper

Surprising findings

• End-to-end training of the full model yields poor PILGD quality and worse performance than the alternate training scheme (44.0 vs 36.4 mm MH protocol 1).

Limitations

• Single-hypothesis performance is comparable to, not clearly better than, state-of-the-art deterministic methods on H3.6M.
• Evaluation limited to two datasets (H3.6M and MPI-INF-3DHP) using 2D ground-truth or detected inputs as described.
• End-to-end training is unstable, requiring a custom alternating training scheme.

Experimental setup

Datasets: Human3.6M

Benchmarks: Human3.6M MPI-INF-3DHP

Metrics: MPJPE PCK AUC

Baselines: Martinez et al. 2017 Li and Lee 2019 Li and Lee 2020 Sharma et al. 2019 Wehrbein et al. 2021 Ci et al. 2023 Ci et al. 2019 Pavllo et al. 2019 Li et al. 2023

Builds on: normalizing flow (RealNVP) Wehrbein et al. 2021 normalizing flow HPE Ci et al. 2023 generative HPE energy-based adversarial training (Zhao et al. 2017) Brehmer and Cranmer 2020 alternate training

Closest papers by meaning

• Normalizing Flows on the Product Space of SO(3) Manifolds for Probabilistic Human Pose Modeling (CVPR 2024 · similarity 0.87)
• Toward Approaches to Scalability in 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.87)
• PoseIRM: Enhance 3D Human Pose Estimation on Unseen Camera Settings via Invariant Risk Minimization (CVPR 2024 · similarity 0.87)
• Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation (AAAI 2024 · similarity 0.86)
• PandaPose: 3D Human Pose Lifting from a Single Image via Propagating 2D Pose Prior to 3D Anchor Space (NeurIPS 2025 · similarity 0.86)
• Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.85)