SoroushMehraban/SynthGait-19K

SynthGait-19K

<div align="center"> <video autoplay loop muted controls playsinline style="max-width: 100%;"> <source src="https://huggingface.co/datasets/SoroushMehraban/SynthGait-19K/resolve/main/teaser.mp4" type="video/mp4"> </video> </div>

SynthGait-19K contains 19,272 synthetic walking videos from 6,427 underlying motion sequences. Each motion is rendered from up to three camera viewpoints. The dataset supports video-based gait parameter estimation.

Dataset structure

The release uses uncompressed WebDataset TAR shards of approximately 1 GB. The shared vid_XXXXX base identifies one walking sequence and the suffix identifies its camera view. Each full stem forms one WebDataset sample:

vid_00001_frontal.mp4   # H.264 video, 832x480, 16 FPS
vid_00001_frontal.npz   # sanitized six-target numeric annotation
vid_00001_frontal.json  # viewer-friendly metadata
vid_00001_saggital.mp4
vid_00001_saggital.npz
vid_00001_saggital.json
vid_00001_random.mp4
vid_00001_random.npz
vid_00001_random.json

Related views are assigned to the same split. View totals are frontal 6,425, sagittal 6,426, and random 6,421.

The JSON member contains sample_id, split, view, video properties, camera rotation and field of view, named gait features, and left/right heel-strike indices. The NPZ contains the released features, left_heel_strikes, and right_heel_strikes arrays.

Heel-strike indices refer to the 30 Hz annotation timeline, not directly to the 16 FPS video frame indices.

Gait feature schema

The features NPZ vector and the gait_features JSON object use this order: cadence, walking_speed, step_length_mean, step_width_mean, stoop_posture_mean, arm_swing. These are the six output targets in the released GaitXFormer configuration. All other legacy source features are excluded from both formats.

Download and build the training cache

The repository's cache builder consumes these shards directly. After cloning the FoundationGait repository and installing its requirements:

export SYNTHGAIT_ROOT="/path/to/SynthGait-19K"
export SYNTHGAIT_CACHE="/path/to/SynthGait-19K-cache"

hf download YOUR_NAMESPACE/SynthGait-19K \
  --repo-type dataset \
  --local-dir "$SYNTHGAIT_ROOT"

python cache_hf_dataset.py \
  --dataset-dir "$SYNTHGAIT_ROOT" \
  --output-dir "$SYNTHGAIT_CACHE" \
  --num-workers 8

The shard name determines the split: train-*.tar becomes cache split train, and validation-*.tar becomes cache split val. The split is finalized during packaging, with every view of an underlying walk kept in the same split; clients do not need a separate filename manifest.

cache_hf_dataset.py reads the six released targets by name and uses a person detector to construct the cropped 32-frame tensors expected by the training loader. It does not require the excluded legacy gait targets.

Public identifiers and metadata

Walks are shuffled before public vid_XXXXX identifiers are assigned. Camera views of the same motion share that base identifier, so no separate walk_id is needed. The release does not include original filenames, subject identifiers, source-dataset labels, or the private dataset_mapping.json file. Train versus validation membership is intentionally visible in the shard path and JSON split field.

These measures remove direct source identifiers from the release, but they should be described as pseudonymization rather than a guarantee of irreversible anonymity. An organization that already possesses the source motions could potentially match records using motion or annotation similarity.

Other loading options

With Hugging Face Datasets:

from datasets import load_dataset

dataset = load_dataset("YOUR_NAMESPACE/SynthGait-19K", streaming=True)
sample = next(iter(dataset["train"]))

With WebDataset:

import webdataset as wds

url = "https://huggingface.co/datasets/YOUR_NAMESPACE/SynthGait-19K/resolve/main/data/train-{00000..00100}-of-00101.tar"
dataset = wds.WebDataset(url, shardshuffle=True).decode().to_tuple("mp4", "npz", "json")

License

SynthGait-19K is available under the SynthGait-19K Research Use License. It may be used for non-commercial research only. Commercial, production, clinical, diagnostic, and patient-care uses are prohibited without prior written permission from the SynthGait-19K authors.

Synthetic data can reproduce biases or artifacts from its motion sources and generation pipeline. This dataset is not clinically validated and is not a substitute for clinical data or professional medical judgment.