Reproduce “BAL: Bidirectional Autoregression in Latent Space for Learning Human Movement Representations” 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

Genki Kinoshita, Ko Nishino

BAL learns pose representations by predicting chunk representations in both temporal directions, using natural sequence order rather than exact future poses. It shares an extractor across directions and uses four-second chunks. On Nymeria captioning, its BERTScore is 29.6 versus 28.5 for MacDiff. Across three downstream tasks, BAL leads on most reported metrics, while chunk duration remains a design choice.

Summary by alphaXiv.

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