[Paper Review] Anti-Confusing: Region-Aware Network for Human Pose Estimation

Xuan Cao, Yanhao Ge+6 more|arXiv (Cornell University)|May 3, 2019

Human Pose and Action Recognition33 references4 citations

TL;DR

This paper proposes Region-Aware Network (RANet), a novel framework for human pose estimation that enhances robustness to confusing textures—such as heavy occlusion, nearby persons, and symmetric body parts—through parsing-based data augmentation, a feature pyramid stem for improved low-level features, effective region extraction for target-specific features, and cascade voting fusion to refine predictions. RANet achieves state-of-the-art performance on MPII and LSP benchmarks, with significant gains on hard-to-predict joints like wrists, ankles, and elbows.

ABSTRACT

In this work, we propose a novel framework named Region-Aware Network (RANet), which learns the ability of anti-confusing in case of heavy occlusion, nearby person and symmetric appearance, for human pose estimation. Specifically, the proposed method addresses three key aspects, i.e., data augmentation, feature learning and prediction fusion, respectively. First, we propose Parsing-based Data Augmentation (PDA) to generate abundant data that synthesizes confusing textures. Second, we not only propose a Feature Pyramid Stem (FPS) to learn stronger low-level features in lower stage; but also incorporate an Effective Region Extraction (ERE) module to excavate better target-specific features. Third, we introduce Cascade Voting Fusion (CVF) to explicitly exclude the inferior predictions and fuse the rest effective predictions for the final pose estimation. Extensive experimental results on two popular benchmarks, i.e. MPII and LSP, demonstrate the effectiveness of our method against the state-of-the-art competitors. Especially on easily-confusable joints, our method makes significant improvement.

Motivation & Objective

• To address the challenge of confusing textures in human pose estimation, including heavy occlusion, nearby persons, and symmetric body appearances.
• To identify symmetric appearance as a previously underexplored source of confusion in pose estimation.
• To develop a unified framework that simultaneously improves data augmentation, feature learning, and prediction fusion for anti-confusing capability.
• To enhance performance on hard-to-predict joints such as wrists, ankles, knees, and elbows.

Proposed method

• Proposes Parsing-based Data Augmentation (PDA) to synthesize realistic confusing textures by segmenting and repositioning body parts on or around human joints.
• Introduces a Feature Pyramid Stem (FPS) that learns multi-scale low-level features across different resolutions to strengthen feature representation in early network stages.
• Designs an Effective Region Extraction (ERE) module that crops human body regions based on intermediate bounding box predictions, reducing background interference and enhancing target-specific features.
• Develops a Cascade Voting Fusion (CVF) module that excludes low-confidence predictions and adaptively fuses high-quality candidates using weighted averaging for final output.
• Integrates all components into a stacked hourglass-based architecture, enabling end-to-end training and inference.

Experimental results

Research questions

• RQ1 Can parsing-based data augmentation effectively simulate complex confusing textures such as occlusion and symmetric body parts?
• RQ2 Does enhancing low-level feature learning through a feature pyramid stem improve pose estimation robustness?
• RQ3 Can effective region extraction from predicted bounding boxes reduce background noise and improve feature specificity?
• RQ4 Does cascade voting fusion that excludes inferior predictions lead to more accurate final pose estimation?

Key findings

• RANet achieves a PCKh@0.5 score of 91.52% on the MPII validation set, outperforming prior state-of-the-art methods.
• The ablation study shows that PDA alone improves accuracy by 1.14%, FPS by 0.35%, ERE by 0.23%, and CVF by up to 1.06% when combined with ERE.
• On hard joints like wrists and ankles, RANet shows significant improvement over baseline and prior methods, with reduced mean and variance of prediction errors.
• The optimal threshold for CVF is near the mean distance of candidates to their center; deviations reduce performance, confirming the importance of selective fusion.
• Statistical analysis confirms that limb joints (especially wrists and ankles) have higher error variance than torso joints, and horizontal prediction is generally more challenging than vertical.