1 of 1

Pose Prior Learner: Unsupervised Categorical Prior Learning for Pose Estimation

Ziyu Wang¹, Shuangpeng Han¹, Mengmi Zhang^1,*

¹Deep NeuroCognition Lab, Nanyang Technological University, Singapore

^*Address correspondence to mengmi.zhang@ntu.edu.sg

Introduction

Pose Prior Learner (PPL)

Challenge

Conclusion

Experimental Results

(a) Learn a category-level pose prior in a fully self-supervised manner.

(b) Enable robust inference under occlusion.

PPL Architecture

Iterative Inference

Introduce unsupervised categorical prior learning as a key challenge in pose estimation
Propose Pose Prior Learner (PPL) to learn pose prior for pose estimation
Outperform existing baselines on pose estimation benchmarks
Enable iterative inference and robust pose estimation under occlusion
Reveal a new perspective on prior learning by distilling shared structure from instance-level predictions

Motivation: Pose estimation benefits from pose priors but obtaining them usually requires annotations.

Problem: Learn a category-level pose prior from images in a self-supervised way.

Method: PPL learns pose priors via hierarchical memory of prototypical poses, improving robustness (e.g., under occlusion).

Quantitative Results

Visualization on Human3.6m

PPL across Categories

Funding sources:

Scan QR code for paper, code, data

(a) PPL outperforms all existing baselines.

(b) PPL refines pose estimation under occlusion with learned pose prior.