SynthGait-19K: A Physically Grounded Synthetic Video Dataset for Gait Parameter Estimation
SynthGait-19K pairs 19,272 synthetic walking videos with motion and six gait-parameter annotations for monocular video research.
The dataset is built from 6,427 MoCap sequences across 437 subjects, converted through SMPL and rendered as controllable RGB walking videos. The authors introduce Gait2Vid for generation and GaitXFormer as a direct-RGB reference model. They report that synthetic supervision transfers to real videos for both GaitXFormer and a pose-based model, while spatial gait parameters remain more exposed to domain shift. HF Daily Papers' note
The dataset is built from 6,427 MoCap sequences across 437 subjects, converted through SMPL and rendered as controllable RGB walking videos. The authors introduce Gait2Vid for generation and GaitXFormer as a direct-RGB reference model. They report that synthetic supervision transfers to real videos for both GaitXFormer and a pose-based model, while spatial gait parameters remain more exposed to domain shift. HF Daily Papers' note
score 5