Balazia2025VUCHA

Title

Video Understanding of Complex Human Activities
View PDF | Save PDF

Authors

Michal Balazia

Affiliations

INRIA Université Côte d’Azur, Sophia Antipolis, France.

Abstract

This report presents a comprehensive collection of novel methods and datasets advancing multimodal AI for behavioral analysis in social interactions, medical training, emotion recognition, and psychiatric phenotyping. Contributions include the MultiMediate'25 challenge for cross-cultural engagement estimation using expanded NOXI and NOXI-J corpora, a Go-ELAN YOLOv9 model for real-time surgical instrument detection in cataract videos, the CM3T adapter framework for efficient multimodal transfer learning on inhomogeneous datasets, the BLEMORE dataset with relative salience annotations for blended emotions, and MEPHESTO analyses revealing context-aware synchrony for therapeutic alliance, temporal variability for depression-schizophrenia classification, and trauma-modulated speech patterns in depression via MADRS/BDI-II assessments. These works of the INRIA-STARS team on video understanding of complex human activities bridge gaps in multimodal behavioral AI, supporting applications from assistive systems to personalized psychiatry.

KeyPhrases

Video understanding, pattern recognition, human behavior analysis, social interactions, psychiatric phenotyping.

Dates

Created 2025-10-02, presented 2025-10-09, updated 2025-12- 13, published 2025-12-13.

Citation

Brainiacs Journal 2025 Volume 6 Issue 3 Edoc F3DAD390E
DOI: 10.48085/F3DAD390E
NPDS: LINKS/Brainiacs/Balazia2025VUCHA