ALE : active learning extension for object detection
Theo Oriol  1, 2@  , Jerôme Pasquet  3, 4, 5@  , Jérome Cortet  6@  
1 : Centre d'Ecologie Fonctionnelle et Evolutive
PSL Research University, CEFE, UMR 5175, CNRS, Université de Montpellier, University Paul-Valéry Montpellier, EPHE
2 : APPLICATION DES MATHÉMATIQUES, INFORMATIQUE ET STATISTIQUE
AMIS, Université de Montpellier 3, Montpellier, France
3 : Territoires, Environnement, Télédétection et Information Spatiale  (UMR TETIS)
Centre de Coopération Internationale en Recherche Agronomique pour le Développement, AgroParisTech, Centre National de la Recherche Scientifique, Institut National de Recherche pour l’Agriculture, l’Alimentation et l’Environnement
4 : Université Paul-Valéry Montpellier 3 - Faculté Éducation et sciences pour les LLASHS  (UPVM UM3 UFR6)
Université Paul-Valéry - Montpellier 3
5 : APPLICATION DES MATHÉMATIQUES, INFORMATIQUE ET STATISTIQUE
AMIS, Université de Montpellier 3, Montpellier, France
6 : Centre d'Ecologie Fonctionnelle et Evolutive
Université Paul-Valéry Montpellier 3, Université de Montpellier, CNRS, EPHE, IRD

Monitoring human activities' impact on soil biodiversity over time is a costly and resource-intensive challenge. Modern technologies like deep learning offer a promising solution due to their capability to analyze large datasets much faster than humans. However, deep learning relies on extensively annotated datasets, and annotating these samples is both time-consuming and expensive, complicating its application. This paper introduces a novel active learning approach called Active Learning Extension (ALE), which aims at improving model performance in object detection tasks while minimizing the need for extensive data annotation. Traditional active learning methods typically rely solely on prediction uncertainty to select images for annotation, which can be suboptimal when introducing new classes. ALE addresses this limitation by considering both the uncertainty and the number of predictions. This dual consideration leads to significant improvements, particularly in scenarios like Collembola detection, where creating and updating datasets is highly timeintensive. Our evaluation demonstrates that ALE significantly enhances model performance compared to stateof-the-art methods. The results underscore the importance of selecting challenging examples and accounting for the number of predictions to optimize active learning in object detection.


Personnes connectées : 1 Vie privée
Chargement...