Communication Dans Un Congrès Année : 2021

Deep Fisher Score Representation via Sparse Coding

Sixiang Xu
  • Fonction : Auteur
  • PersonId : 1149811
  • IdRef : 263577163
Damien Muselet
Alain Trémeau
  • Fonction : Auteur
  • PersonId : 859601

Résumé

Fisher Score has been shown to be accurate global image features for classification. Most of time, it is based on a Gaussian mixture model (GMM). Nevertheless, recent studies show that GMM does not fit well high dimensional data such as the ones extracted by deep convolutional networks. In this paper, we propose to resort to a sparse representation of the centers of the Gaussian functions in order to better cover the high dimensional feature space. This solution has already been used in a framework constituted by independent and off-the-shelf modules and the contribution of this paper is to embed these steps in an end-to-end deep neural network so that all the modules work together for the sole purpose of improving classification performance. Experimental results show that this solution clearly outperforms many alternatives in the context of material, indoor scenes or fine-grained image classification.
Fichier principal
Vignette du fichier
CAIP21.pdf (421.47 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

ujm-03726540 , version 1 (30-01-2025)

Identifiants

Citer

Sixiang Xu, Damien Muselet, Alain Trémeau. Deep Fisher Score Representation via Sparse Coding. International Conference on Computer Analysis of Images and Patterns, Sep 2021, Nicosia, Cyprus. pp.412-421, ⟨10.1007/978-3-030-89131-2_38⟩. ⟨ujm-03726540⟩
16 Consultations
0 Téléchargements

Altmetric

Partager

More