SynthGait–19K

A Physically Grounded Synthetic Video Dataset
for Gait Parameter Estimation

Soroush Mehraban^1,2,3 Xin Lei Lin^1,2,3 Vida Adeli^1,2,3 Majid Mirmehdi⁴ Amirhossein Dadashzadeh⁴ Clint Hansen⁵ Andrea Iaboni^1,2 Babak Taati^1,2,3

¹KITE Research Institute, University Health Network (UHN) ²University of Toronto ³Vector Institute ⁴University of Bristol ⁵Kiel University

PaperPDF PaperarXivSoon DownloadDataset Try onlineDemo RepositoryCode

Explore the dataset

Paired by construction

Depth-conditioned video samples.

Move across the frame to reveal the geometric depth conditioning beneath each generated RGB video. On touch screens, drag anywhere on the frame.

Depth RGB

Sample 01

Office walk

Front / back view

Hover or drag to compare

19,272RGB videos

6,427MoCap sequences

437Subjects

671Minutes of walking

6Gait parameters

What is SynthGait-19K?

Dataset construction
and gait annotations.

SynthGait-19K is a large-scale synthetic walking-video dataset built from real motion-capture recordings, not arbitrary animation.

Gait2Vid unifies five heterogeneous MoCap datasets through SMPL, places each motion under controllable cameras, and uses depth-conditioned video diffusion to create diverse people and environments while retaining correspondence with the source motion.

Every RGB clip stays paired with its fitted 3D SMPL motion and annotations for six clinically meaningful gait parameters.

Cadence

Heel-strike timing

Walking speed

Pelvis displacement

Step length

Forward foot travel

Step width

Mediolateral spacing

Stooped posture

Neck–pelvis offset

Arm swing

Wrist motion range

From capture to benchmark

Controlled viewpoint and appearance variation.

The pipeline preserves the biomechanics of recorded walking while varying camera geometry, visual appearance, and scene context independently.

Figure 02

The Gait2Vid pipeline

Unify heterogeneous MoCap as SMPL motion.
Observe each walk from front, side, and oblique cameras.
Synthesize RGB with depth and scene prompts conditioning VACE.
Annotate gait directly from the fitted motion.

Reference model

Direct gait-parameter estimation from RGB video.

A V-JEPA2-initialized Video ViT encodes a walking clip into spatiotemporal tokens. Six learnable gait queries then cross-attend to that video representation—one dedicated query for each target quantity.

0.84Average correlation
on real GPJATK

0.27sPer 5-second clip
on an RTX 3090

Figure 04 GaitXFormer architecture

Dataset composition

Motion sources and cohort composition.

Five public MoCap sources contribute different populations and protocols. Hover or focus the composition below to inspect each source.

Video composition19,272 total

View full composition table +

Motion source	Subjects	Sequences	Front / back	Sagittal	Oblique	Total videos
Schreiber & Moissenet	49	917	917	918	918	2,752
Santos et al.	25	488	488	488	488	1,464
Bertaux et al.	186	4,442	4,441	4,442	4,436	13,319
Grouvel et al.	10	82	82	82	82	246
Warmerdam et al.	167	497	497	497	497	1,491
SynthGait-19K total	437	6,427	6,425	6,427	6,421	19,272

Table 04 · Real-world benchmark

Comparison on real GPJATK video.

Trained with SynthGait-19K supervision and evaluated on real GPJATK video, GaitXFormer obtains the highest Fisher-averaged correlation among the compared methods.

Average 0.84 Pearson r

Preferred-view Pearson correlation on real GPJATK. Higher is better. Bold values mark the best result for each gait parameter.
Method	CadenceCad.	Walking speedW. speed	Step lengthStep len.	Step widthStep wid.	Stooped postureStoop post.	Arm swingArm swing	Average
Human mesh recovery
WHAM	0.85	0.82	0.40	0.69	0.64	0.67	0.70
CameraHMR	0.71	0.65	0.10	0.35	0.39	0.91	0.59
PromptHMR	0.87	0.61	0.13	0.48	0.63	0.90	0.67
FastHMR	0.80	0.63	−0.07	0.18	0.02	0.93	0.54
Biomechanical
PBL	0.47	0.77	0.38	0.50	0.73	0.63	0.60
OpenCap-M	0.76	0.87	0.62	0.47	0.58	0.65	0.68
2D pose-based
STT	0.20	0.67	—	—	—	—	—
STT^†SynthGait trained	0.91	0.85	0.73	0.67	0.70	0.92	0.82
Direct video-based
GaitXFormerSynthGait-19K trainedProposed	0.94	0.88	0.67	0.67	0.75	0.91	0.84

Synthetic-to-real evaluation. GaitXFormer obtains an average correlation of 0.84, compared with 0.82 for the next-highest method and 0.70 for the highest-scoring HMR pipeline.

Cell shading encodes correlation strength

Training-set scale analysis

Performance by training-set size.

GaitXFormer’s average Pearson correlation on real GPJATK rises as the SynthGait training subset grows. Select a training fraction to inspect every parameter.

Fisher-averaged correlation0.84

Cadence0.94

Walking speed0.88

Step length0.67

Step width0.67

Stooped posture0.75

Arm swing0.91

0.0Pearson correlation (r)1.0

Read the paper