VClipper: Exploiting CLIP Zero-shot capabilities for moment retrieval in video recordings

This research explores the integration of CLIP, a pretrained model, into video content analysis. In a landscape inundated with multimedia data, pinpointing specific moments within videos is a persistent challenge. By leveraging CLIP's semantic and visual search capabilities, this study endeavor...

Descripción completa

Detalles Bibliográficos
Autor: Caravaca Müller, Oriol
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Oberta de Catalunya (UOC)
Repositorio:O2, repositorio institucional de la UOC
OAI Identifier:oai:openaccess.uoc.edu:10609/149808
Acceso en línea:http://hdl.handle.net/10609/149808
Access Level:acceso abierto
Palabra clave:video analysis
moment retrieval
CLIP
Computer vision -- TFM
Visió per ordinador -- TFM
Descripción
Sumario:This research explores the integration of CLIP, a pretrained model, into video content analysis. In a landscape inundated with multimedia data, pinpointing specific moments within videos is a persistent challenge. By leveraging CLIP's semantic and visual search capabilities, this study endeavors to refine content retrieval methods. Emphasizing efficiency and applicability, this study aims to make this process more precise and practical. With this research we also reviewed the state-of-the-art methods and produced empirical analysis on the effects of postprocessing on the similarity vectors obtained from CLIP encoders. Finally, we developed two distinct methods aimed at moment retrieval tasks in audiovisual data, obtaining a model that is able to outperform previous works in Zero-shot moment revival, reaching 57.3 at R@1 IoU=0.5 and 51.6 at mAP@0.5.