Neural Human Pose Prior

Jul 16, 2026

Upload by Himbo

A principled, data-driven neural prior for human body poses is presented using normalizing flows. The approach learns a flexible pose density over parameters expressed in a 6D rotation format, avoiding heuristic or low-expressivity alternatives. The key challenge of modeling distributions on the manifold of valid 6D rotations is handled by inverting the Gram-Schmidt process during training, yielding stable learning while remaining compatible with rotation-based pipelines. A reproducible, framework-agnostic architecture supports qualitative and quantitative evaluation plus ablation analysis, enabling probabilistic pose priors for motion capture and reconstruction.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

Introduction

Background

Losses vs. Probabilities

Losses

Unidentifiable Problems

Bayesian Statistics

Neural Human Pose Prior Michal Heker Yoom [michal@yoom.com](michal@yoom.com) Sefy Kagarlitsky Yoom [sefy@yoom.com](sefy@yoom.com) David Tolpin Yoom [davidt@yoom.com](davidt@yoom.com) arXiv :2507 . 12138v2 [ cs .CV] 11 Jul 2026 July 14, 2026 Abstract We introduce a principled, data-driven approach for modeling a neural prior over human body poses using normalizing flows. Unlike heuristic or low-expressivity alternatives, our method leverages RealNVP to learn a flexible density over poses represented in the 6D rotation format. We address the challenge of modeling distributions on the manifold of valid 6D rotations by inverting the Gram-Schmidt process during training, enabling stable learning while preserving downstream compatibility with rotationbased frameworks. Our architecture and training pipeline are frameworkagnostic and easily reproducible. We demonstrate the effectiveness of the learned prior through both qualitative and quantitative evaluations, and we analyze its impact via ablation studies. This work provides a sound probabilistic foundation for integrating pose priors into human motion capture and reconstruction pipelines. 1 Introduction Human motion capture is concerned with obtaining a parametric representation of human motion from raw inputs such as single-view or multiview videos. Even in a studio setup, the inputs are often noisy and partial—consider, for example, a group of people moving together, casting shadows and occluding each other relative to the camera. A common consensus is that some population-based bias — a prior — should be introduced into the model for more reliable motion capture. However, constructing or learning a human motion prior is too challenging due to both scarcity of data and complexity and diversity of human motions. A viable compromise though is to come up with a human pose prior. A parametric human body model commonly has only a few dozen joints with at most three degrees of rotational freedom each, a dimensionality well within the capabilities of modern machine learning approaches. Intuitively, a prior should be a density function over human pose parameter vectors such that the gradient of the function points in the direction of the quickest ‘improvement’ of the pose. In addition, since human poses lie on a low-dimensional manifold embedded into the Euclidean space of pose parameters, the ambient pose space should be projected into a latent space in which sufficiently small Euclidean vicinity of a viable pose consists of poses that are only slightly more or less viable. These two constituents—reparameterization and a density function—are the essentials of a body pose prior. Despite the apparent simplicity of the task, it turns out that coming up with a coherent, scientifically justified procedure of constructing a human body prior is quite challenging and hard to get right (Section 3) . In this work, we establish a clean, principled scheme for training and using during inference a human pose prior. Our contributions are as follows: • We establish a systematic foundation of a data-driven human pose prior. • We describe, in a framework-agnostic way and in detail sufficient for easy reproduction, our network architecture and training routine. • We evaluate our approach empirically, quantitatively and qualitatively, compare with baselines, and support our design choices by ablation studies. 2 Background 2. 1 Losses vs. Probabilities 2.1.1 Losses Consider a task of finding a number the square of which is 9 . x2 = 9 An obvious solution is the closed-form one. If f(x) = x2 then f−1(y) = √y , so if x2 = 9 then x = √9 = 3 . However, we do not always have f −1 in closed form. An alternative is to use a numerical method of sort, for example, gradient descent: • Parameter: x • Ground truth: 9 • Transformation: x2 • Loss (L2 ): (9 − x2 )2 ∇x (9 − x2 )2 = −4(9 − x2 )x If we make a small step in the direction of 4(9 − x2 )x, usually pro

FAQs

How does the method model a pose prior over human poses?

It uses normalizing flows to learn a neural probability density over pose parameters, represented in a 6D rotation format.

What problem arises when learning distributions over 6D rotations, and how is it addressed?

Valid 6D rotations lie on a manifold, so the method inverts the Gram-Schmidt process during training to make manifold-constrained learning stable and compatible with downstream rotation frameworks.

How is the proposed pose prior validated in the work?

Effectiveness is shown through qualitative and quantitative evaluations, supported by ablation studies analyzing the impact of design choices.

Jul 16, 2026

Upload by Himbo