Automated mask generation for efficient YOLO-based instance segmentation in marine environments for fish detection

This paper addresses the laborious, time-consuming and error- prone process of generating ground truth data to perform instance segmentation of fish in their natural habitat. Our proposal is to use the Segment Anything Model (SAM), which allows zero- shot inference, to automatically build the segmen...

ver descrição completa

Detalhes bibliográficos
Autores: Rovira Coll, Xènia, Burguera Burguera, Antoni
Formato: artículo
Fecha de publicación:2024
País:España
Recursos:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/411427
Acesso em linha:https://hdl.handle.net/2117/411427
https://dx.doi.org/10.5821/iwp.2024.23.14134
Access Level:acceso abierto
Palavra-chave:Animals -- Identification
Machine learning
Instance segmentation
Deep learning
SAM
YOLO
Animals -- Identificació
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Descrição
Resumo:This paper addresses the laborious, time-consuming and error- prone process of generating ground truth data to perform instance segmentation of fish in their natural habitat. Our proposal is to use the Segment Anything Model (SAM), which allows zero- shot inference, to automatically build the segmentation masks, significantly reducing the dataset creation time and enhancing scalability to larger datasets. Experimental results using You Only Look Once (YOLO) demonstrate only marginal performance differences between our approach –with segmentation masks created with no human intervention– and a standard training using a fully human-labeled dataset. The results underscore the effectiveness of the automated workflow discussed herein, showcasing substantial reduction in dataset creation time, particularly in demanding underwater scenarios.