Reproduce “SMPLer-X: Scaling Up Expressive Human Pose and Shape Estimation” 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
Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Yanjun Wang, Hui En Pang, Haiyi Mei and 5 more
Researchers from S-Lab, Nanyang Technological University, SenseTime Research, and other institutions developed SMPLer-X, a generalist foundation model for expressive human pose and shape estimation (EHPS). This model is trained on up to 32 diverse EHPS datasets, totaling 4.5 million instances, and utilizes large Vision Transformer backbones, achieving state-of-the-art performance with a PVE of 69.5 mm on AGORA and an overall mean primary error (MPE) of 63.1 mm for its largest configuration.
Key insights
- Models trained on individual datasets showed strong in-domain performance but significant degradation on different domains, underscoring the necessity of diverse, large-scale data for generalizability.
- Both increasing data diversity (number of distinct datasets, scene types, real/synthetic appearance) and model size (larger Vision Transformer backbones) systematically improved generalizability and overall EHPS performance.
- Finetuning the generalist foundation model on specific target training sets, even by only adapting neck and head modules, provides substantial performance boosts for domain-specific tasks and offers an efficient adaptation strategy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.