Gaze Target Detection with Diffusion Models

Gaze target detection aims to identify where an individual is looking within an image. This task is critical for many downstream tasks, such as human-robot interaction, cognitive load estimation, and human attention estimation. This thesis addresses the task of gaze target detection, exploring exist...

Descripción completa

Detalles Bibliográficos
Autor: Amenábar Fábregas, Núria
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/422214
Acceso en línea:https://hdl.handle.net/2117/422214
Access Level:acceso abierto
Palabra clave:Image processing
Deep learning (Machine learning)
Diffusion models
Gaze Target Detection
Gaze following
Gaze estimation
deep learning
image generation
saliency map
Imatges--Processament
Aprenentatge profund
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Descripción
Sumario:Gaze target detection aims to identify where an individual is looking within an image. This task is critical for many downstream tasks, such as human-robot interaction, cognitive load estimation, and human attention estimation. This thesis addresses the task of gaze target detection, exploring existing models and introducing a novel approach utilizing diffusion models. Initially, an existing transformer-based model for gaze target detection has been adapted to be trained and tested on a new dataset. Subsequently, a new architecture based on diffusion models is proposed, incorporating a diffusion process conditioned by RGB image information and bounding box locations of the individuals. The objective is to generate a gaze target saliency map, indicating where individuals marked by the bounding boxes are focusing their attention. As the core section of the research, the exploration of the joint attention estimation task is performed using manually annotated bounding boxes, followed by experimentation using head bounding boxes detected by an object detection neural network, as well as the task of estimating single attention. Through extensive analysis of the model on the VideoCoAtt dataset against the baseline transformer-based model, although the metric does not meet the desired threshold for improvement, the visual outputs demonstrate promising results. This suggests potential avenues for further refinement and optimisation of diffusion models applied to the task of gaze target detection.