SynthGait-19K: 19,272 synthetic clips train monocular gait models

A new arxiv preprint introduces SynthGait-19K, a physically grounded synthetic video dataset of 19,272 walking clips derived from 6,427 motion-capture sequences across 437 subjects, released with paired SMPL motion and annotations for six gait parameters. It targets a well-known bottleneck in clinical mobility AI: existing real-world gait datasets are small, narrow in viewpoint, and visually thin, which caps how well monocular estimators generalize.

The team's Gait2Vid pipeline "unifies heterogeneous MoCap recordings through SMPL and synthesizes diverse RGB walking videos under controllable viewpoints and scene appearances." Extracted gait events were validated against force-platform measurements, and the generated clips were checked for consistency with their conditioning kinematics. Alongside the dataset the authors introduce GaitXFormer as a direct RGB reference model, and benchmark it against pose-based, biomechanical, and human-mesh-recovery approaches.

The headline result is that "synthetic supervision transfers effectively to real videos across both GaitXFormer and a pose-based architecture." The abstract also flags where the transfer weakens: "spatial gait parameters are more sensitive to visual domain shift," and "improved HMR reconstruction alone does not necessarily translate to improved downstream gait estimation."

The abstract does not name the six parameters, publish per-parameter accuracy numbers, or describe any patient population used for real-video evaluation.