StimBench: A Video Benchmark for Stereotypical Motor Movement Detection in Autism Spectrum Disorder StimBench: OSB'de Stereotipik Hareketlerin Tespiti için Video Tabanli Ölçüt Veri Kümesi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636777
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: autism spectrum disorder, parameter-efficient fine-tuning, stereotypical behavior detection, V-JEPA 2, video classification
- Gazi Üniversitesi Adresli: Evet
Özet
Existing video datasets for stereotypical motor movement (stimming) detection in Autism Spectrum Disorder are distributed as YouTube URL lists that degrade over time, limiting reproducibility. We introduce StimBench, a frozen, face-anonymized benchmark of 333 clips across three stimming classes and a normal class, with video-level splits and a gender-balanced test set. Across 27 configurations spanning six architectures and four parameter-efficient fine-tuning (PEFT) strategies, V-JEPA 2 with LoRA achieves the best result at over 90 percent accuracy while updating only a small fraction of parameters. PEFT consistently outperforms full fine-tuning under limited clinical data. Gender-stratified analysis shows the largest male-female gaps among fully fine-tuned models (e.g., I3D at 57.1%/41.2%), though gaps remain variable across adapted models as well. The dataset and code are publicly available at https://github.com/UlkuTuncerKucuktas/StimBench.