Reproduce “Regressive Domain Adaptation for Unsupervised Keypoint Detection” 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
Junguang Jiang, Yifei Ji, Ximei Wang, Yufeng Liu, Jianmin Wang, Mingsheng Long
Regressive Domain Adaptation (RegDA) enables unsupervised keypoint detection by re-engineering adversarial training for continuous output spaces, converting the minimax problem into two complementary minimization goals. This method yielded up to a 10.7% PCK improvement on real-world hand and human pose estimation tasks, transferring knowledge from synthetic to unlabeled real images.
Key insights
- Traditional adversarial minimax objectives struggle in high-dimensional continuous output spaces, often increasing prediction variance rather than shifting the mean to distinct adversarial locations.
- Incorrect keypoint predictions are not random but tend to cluster around other semantically plausible keypoint locations, a 'sparsity' in the output space that can be leveraged.
- Converting the adversarial minimax problem into two distinct minimization goals, one for guiding the adversarial regressor and another for aligning features, provides a stable and effective strategy for regression domain adaptation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.