Reproduce “Human Pose Estimation with Iterative Error Feedback” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, Jitendra Malik

Iterative Error Feedback (IEF) turns 2D human pose estimation into a sequence of bounded corrections: a ConvNet repeatedly combines the image with heatmaps of its current joint estimate and updates that estimate. On MPII, it reaches 81.3 PCKh-0.5 without ground-truth person scale and 81.3 with scale supplied, while ablations show that feedback, error targets, and progressive training each contribute to the gains.

Key insights

  • Representing an intermediate pose as spatial heatmaps lets ordinary convolutional layers jointly process image evidence and structured output context.
  • Bounded error regression converts a difficult global localization problem into a series of local refinement steps; iteratively predicting full coordinates performs substantially worse.
  • Feedback from related joints improves localization: for left-knee prediction, adding hip information and then all joint channels raises validation PCKh from 64.6 to 73.8.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.