Carregant...
Tipus de document
TesiVersió
Versió publicadaData de publicació
Llicència de publicació
Si us plau utilitzeu sempre aquest identificador per citar o enllaçar aquest document: https://hdl.handle.net/2445/206040
Towards Video Transformers for Automatic Human Analysis
Títol de la revista
Autors
Director/Tutor
ISSN de la revista
Títol del volum
Resum
[eng] With the aim of creating artificial systems capable of mirroring the nuanced understanding and interpretative powers inherent to human cognition, this thesis embarks on an exploration of the intersection between human analysis and Video Transformers. The objective is to harness the potential of Transformers, a promising architectural paradigm, to comprehend the intricacies of human interaction, thus paving the way for the development of empathetic and context-aware intelligent systems. In order to do so, we explore the whole Computer Vision pipeline, from data gathering, to deeply analyzing recent developments, through model design and experimentation.
Central to this study is the creation of UDIVA, an expansive multi-modal, multi-view dataset capturing dyadic face-to-face human interactions. Comprising 147 participants across 188 sessions, UDIVA integrates audio-visual recordings, heart-rate measurements, personality assessments, socio- demographic metadata, and conversational transcripts, establishing itself as the largest dataset for dyadic human interaction analysis up to this date. This dataset provides a rich context for probing the capabilities of Transformers within complex environments. In order to validate its utility, as well as to elucidate Transformers' ability to assimilate diverse contextual cues, we focus on addressing the challenge of personality regression within interaction scenarios. We first adapt an existing Video Transformer to handle multiple contextual sources and conduct rigorous experimentation. We empirically observe a progressive enhancement in model performance as more context is added, reinforcing the potential of Transformers to decode intricate human dynamics. Building upon these findings, the Dyadformer emerges as a novel architecture, adept at long-range modeling of dyadic interactions. By jointly modeling both participants in the interaction, as well as embedding multi- modal integration into the model itself, the Dyadformer surpasses the baseline and other concurrent approaches, underscoring Transformers' aptitude in deciphering multifaceted, noisy, and challenging tasks such as the analysis of human personality in interaction.
Nonetheless, these experiments unveil the ubiquitous challenges when training Transformers, particularly in managing overfitting due to their demand for extensive datasets. Consequently, we conclude this thesis with a comprehensive investigation into Video Transformers, analyzing topics ranging from architectural designs and training strategies, to input embedding and tokenization, traversing through multi-modality and specific applications. Across these, we highlight trends which optimally harness spatio-temporal representations that handle video redundancy and high dimensionality. A culminating performance comparison is conducted in the realm of video action classification, spotlighting strategies that exhibit superior efficacy, even compared to traditional CNN-based methods.
Descripció
Matèries (anglès)
Citació
Col·leccions
Citació
SELVA CASTELLÓ, Javier. Towards Video Transformers for Automatic Human Analysis. [consulta: 5 de desembre de 2025]. [Disponible a: https://hdl.handle.net/2445/206040]