SynthGait-19K: A Physically Grounded Synthetic Video Dataset for Gait Parameter Estimation
Soroush Mehraban, Xin Lei Lin, Vida Adeli, Majid Mirmehdi, Amirhossein Dadashzadeh, Clint Hansen, Andrea Iaboni, Babak Taati
KITE Research Institute · University of Toronto · Vector Institute · University of Bristol · Kiel University
Published September 8, 2026 · on the feed September 9, 2026 · ▲ 17 · arXiv 2609.08108
The abstract
Accurate estimation of clinically meaningful gait parameters from monocular video is important for scalable mobility assessment, yet progress is limited by the small scale, restricted viewpoints, and limited visual diversity of existing datasets. We introduce SynthGait-19k, a physically grounded synthetic video dataset containing 19,272 walking videos derived from 6,427 MoCap sequences across 437 subjects, with paired SMPL motion and annotations for six gait parameters. To construct the dataset, we develop Gait2Vid, which unifies heterogeneous MoCap recordings through SMPL and synthesizes diverse RGB walking videos under controllable viewpoints and scene appearances. We assess the generated videos for consistency with their conditioning gait kinematics and validate extracted gait events against force-platform measurements. Using SynthGait-19K, we benchmark direct RGB, pose-based, biomechanical, and human-mesh-recovery approaches and analyze viewpoint, training-data scale, and synthetic-to-real domain shift. We also introduce GaitXFormer as a direct RGB reference model for estimating gait parameters. Synthetic supervision transfers effectively to real videos across both GaitXFormer and a pose-based architecture, demonstrating utility across different representations. We further find that spatial gait parameters are more sensitive to visual domain shift and that improved HMR reconstruction alone does not necessarily translate to improved downstream gait estimation.
The Brief
2:37A short podcast summary of the paper. Written and voiced by AI.
Transcript
SynthGait-19K. A Physically Grounded Synthetic Video Dataset for Gait Parameter Estimation. By Soroush Mehraban and seven others, from the KITE Research Institute, the University of Toronto, and three other institutions. Published on the eighth of September. Its code repository is public and real: about twenty-five source files.
Estimating clinically meaningful gait parameters from ordinary monocular video would make mobility assessment scalable. Progress has been held back by the datasets: they are small, their viewpoints are restricted, and their visual diversity is limited. SynthGait-19K is a physically grounded synthetic video dataset built to remove that limit. Its answer is synthesis: video that is generated, but grounded in real recorded motion.
It contains nineteen thousand two hundred and seventy-two walking videos, derived from six thousand four hundred and twenty-seven MoCap sequences across four hundred and thirty-seven subjects. Each comes with paired SMPL motion and annotations for six gait parameters. The motion in every video therefore comes from a real recorded person, and the video shows it from a chosen viewpoint in a chosen scene. The authors also provide GaitXFormer, a direct RGB reference model for estimating gait parameters.
The videos are made with Gait2Vid, a pipeline that unifies heterogeneous MoCap recordings through SMPL and then synthesises diverse RGB walking videos with controllable viewpoints and scene appearance. The physical grounding is checked twice. The generated videos are assessed for consistency with the gait kinematics that conditioned them, so the rendered walking matches the recorded motion, and gait events extracted from them are validated against force- platform measurements. That second check ties the synthetic data back to a physical measurement of real walking.
The dataset is for researchers building video-based gait estimation. The authors use it to benchmark direct RGB, pose-based, biomechanical, and human- mesh-recovery approaches, and to analyse the effects of viewpoint, training- data scale, and domain shift from synthetic to real. Synthetic supervision transfers effectively to real videos, both for GaitXFormer and for a pose- based architecture, so its usefulness is not tied to one representation. That transfer is the result the dataset most needs to show, since a synthetic dataset is only useful if what is learned from it holds up on real footage.
Two findings will shape how it gets used. Spatial gait parameters are more sensitive to visual domain shift than others. And better HMR reconstruction on its own does not necessarily mean better gait estimation downstream, so improving the body model is not a shortcut to improving the measurement.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is too few videos of people walking, filmed from too few angles.
You want to measure how someone walks, their cadence, speed and step length, from an ordinary phone video, so mobility checks could happen anywhere instead of in a gait lab.
The road ahead
- 1 · The problem ~6 min Measuring gait from ordinary video needs lots of varied, well-labelled training footage, and the existing gait video datasets are small, narrow and costly to grow.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Generate walking videos whose motion comes from real motion-capture recordings, render each walk many ways with fixed labels, and check the videos stay faithful.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min The Gait2Vid pipeline unifies motion recordings, renders and synthesises videos, checks the videos and the gait events, then the dataset is used to benchmark gait methods on real video.
- 6 · The experiment ~6 min You’ll guess: When human-mesh-recovery methods reconstructed the body more accurately, did their gait parameter estimates improve along with it?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.