StimBench: A Benchmark for Stereotypical Motor Movement Detection in Autism

Overview

StimBench is a curated video clip dataset for detecting stereotypical motor movements (stimming) in children with Autism Spectrum Disorder. It merges and cleans four existing stimming datasets into a single standardized benchmark with proper train/test splits, no data leakage, and face anonymization.

Dataset Statistics

Split	ArmFlapping	HeadBanging	Spinning	Normal	Total
Train	76	31	43	120	270
Test	15	8	10	30	63
Total	91	39	53	150	333

Sources

Stimming clips are derived from four publicly available datasets:

• SSBD (Rajagopalan et al., 2013) — 75 entries
• ESBD (OckerGui, 2022) — 117 entries
• WEI-BD (OckerGui, 2022) — 9 entries
• SSBD+ (SARL-IIITB, 2023) — 46 entries

Normal clips are sourced from Kinetics-400 (Kay et al., 2017), filtered for 23 child-relevant activity classes including crawling baby, clapping, dancing, playing with pets, drawing, riding a bike, and playing basketball.

Classes

Class	Description	Source
ArmFlapping	Repetitive flapping of arms/hands, hand waving, finger play	SSBD/ESBD/SSBD+
HeadBanging	Repetitive head banging or hitting	SSBD/ESBD/SSBD+
Spinning	Whole-body spinning/rotating	SSBD/ESBD/SSBD+
Normal	Non-stimming everyday activities	Kinetics-400

Key Properties

• No data leakage: Split at video level — no source video appears in both train and test
• Gender-balanced test set: Equal male/female representation per stimming category
• Face anonymized: All faces blurred using CenterFace detector (deface, threshold=0.2)
• Normal class from external source: Kinetics-400 clips prevent scene-level shortcut exploitation

File Structure

StimBench/
├── train/
│   ├── armflapping/    001.mp4 ... 076.mp4
│   ├── headbanging/    001.mp4 ... 031.mp4
│   ├── spinning/       001.mp4 ... 043.mp4
│   └── normal/         001.mp4 ... 120.mp4
├── test/
│   ├── armflapping/    001.mp4 ... 015.mp4
│   ├── headbanging/    001.mp4 ... 008.mp4
│   ├── spinning/       001.mp4 ... 010.mp4
│   └── normal/         001.mp4 ... 030.mp4
├── metadata.csv
└── README.md

Loading

from datasets import load_dataset

dataset = load_dataset("videofolder", data_dir="StimBench")

Metadata

Each clip has the following metadata in metadata.csv:

Field	Description
file_name	Relative path (e.g., train/armflapping/001.mp4)
label	Class name
split	train / test
type	stimming / normal
source_dataset	SSBD / ESBD / WEI BD / SSBD+ / kinetics-400
group_id	YouTube video ID (for GroupKFold)
url	Original YouTube URL
clip_start	Start time in source video (seconds)
clip_end	End time in source video (seconds)
clip_duration	Clip length (seconds)
gender	M / F / Unknown (stimming clips only)

License

CC-BY-NC-4.0. Original stimming videos are from YouTube via SSBD/ESBD/SSBD+. Normal videos are from Kinetics-400. This dataset provides curated clips with annotations and standardized splits for research purposes only.