Reproduce “CIGPose: Causal Intervention Graph Neural Network for Whole-Body Pose Estimation” on OpenResearch

About this paper

Bohao Li, Zhicheng Cao, Huixian Li, Yangming Guo

CIGPose estimates whole-body 2D poses across 133 keypoints. It scores uncertain keypoint evidence, replaces selected representations with learned keypoint-specific embeddings, and uses hierarchical graph reasoning over body structure. On COCO-WholeBody, CIGPose-x reports 67.0 AP at 384×288, versus 66.5 for DWPose-l. Removing the intervention module reduces AP to 66.8, indicating a small incremental contribution; uncertainty can also replace valid rare poses.

Key insights

• Visual context acts as a confounder in 2D whole-body pose estimation, creating non-causal 'backdoor paths' that lead models to learn spurious correlations between visual features and pose.
• Predictive uncertainty serves as an effective proxy for identifying confounded keypoint representations, with higher uncertainty correlating with keypoints affected by severe confounders like occlusion.
• Counterfactual replacement, by substituting confounded embeddings with learned context-invariant canonical representations, effectively severs spurious causal paths, enabling models to reason about genuine causal effects.

Summary by alphaXiv.