Self-supervised continual learning for object recognition in image sequences - CNRS - Centre national de la recherche scientifique Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Self-supervised continual learning for object recognition in image sequences

Résumé

The autonomous learning of different objects in images, with a continual and unsupervised context, relies on detecting unknown objects and recognizing known ones based on the learned visual representation. Novelty detection is challenging because of the internal representation drifts of known objects not been seen for a long time. Most existing approaches either perform offline unsupervised learning on a large dataset, or continual supervised learning. Nevertheless, very few existing approaches propose unsupervised continual learning for object recognition. In this paper, we propose a new neural network-based approach for continually learning representations of objects from image sequences, that is able to autonomously detect novel objects and to recognize previously learned ones during training. It is based on a statistical test, performed on internal representations, adapted to counterbalance the concept drift, without storing any image. Experimental results show that our approach outperforms the state of the art on MNIST and Fashion-MNIST datasets. In particular, our approach avoids to over-segment the distribution of clusters, which artificially increases traditional indicators such as clustering accuracy.
Fichier principal
Vignette du fichier
ICONIP_final_submission.pdf (368.91 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03465149 , version 1 (03-12-2021)

Identifiants

  • HAL Id : hal-03465149 , version 1

Citer

Ruiqi Dai, Mathieu Lefort, Frédéric Armetta, Mathieu Guillermin, Stefan Duffner. Self-supervised continual learning for object recognition in image sequences. International Conference on Neural Information Processing (ICONIP), Dec 2021, Bali, Indonesia. ⟨hal-03465149⟩
92 Consultations
202 Téléchargements

Partager

Gmail Facebook X LinkedIn More