In this work we propose a method for object recognition based on a random selection of interest regions, heterogeneous set of texture descriptors and a bag-of-features approach based on several k- means clustering runs for obtaining different codebooks. The proposed system is not based on complex region detection as SIFT but on a simple exhaustive extraction of sub-windows of a given image. In the classification step an ensemble of random subspace of support vector machine (SVM) is used. The use of random subspace ensemble coupled to the principal component analysis for reducing the dimensionality of the descriptors permits to reduce the curse of dimensionality problem. In the experimental section we show that the combination of classifiers trained using different descriptors permits a consistent improvement of the performance of the stand alone approaches. The proposed system has been tested on four datasets: in the VOC2006 dataset, in a wide-used scene recognition dataset, in the well-known Caltech-256 Object Category Dataset and in a landmark dataset, obtaining remarkable results with respect to other state-of-the-art approaches. The MATLAB code of our system is publicly available.

Heterogeneous bag-of-features for object/scene recognition

LUMINI, ALESSANDRA
2013

Abstract

In this work we propose a method for object recognition based on a random selection of interest regions, heterogeneous set of texture descriptors and a bag-of-features approach based on several k- means clustering runs for obtaining different codebooks. The proposed system is not based on complex region detection as SIFT but on a simple exhaustive extraction of sub-windows of a given image. In the classification step an ensemble of random subspace of support vector machine (SVM) is used. The use of random subspace ensemble coupled to the principal component analysis for reducing the dimensionality of the descriptors permits to reduce the curse of dimensionality problem. In the experimental section we show that the combination of classifiers trained using different descriptors permits a consistent improvement of the performance of the stand alone approaches. The proposed system has been tested on four datasets: in the VOC2006 dataset, in a wide-used scene recognition dataset, in the well-known Caltech-256 Object Category Dataset and in a landmark dataset, obtaining remarkable results with respect to other state-of-the-art approaches. The MATLAB code of our system is publicly available.
Loris Nanni;Alessandra Lumini
File in questo prodotto:
Eventuali allegati, non sono esposti

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11585/153541
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 27
  • ???jsp.display-item.citation.isi??? 21
social impact