Skip to Main content Skip to Navigation

Reconnaissance de catégories d'objets et d'instances d'objets à l'aide de représentations locales

Eric Nowak 1
1 LEAR - Learning and recognition in vision
Grenoble INP - Institut polytechnique de Grenoble - Grenoble Institute of Technology, LJK - Laboratoire Jean Kuntzmann, Inria Grenoble - Rhône-Alpes
Abstract : Object recognition is one of the most active fields of computer vision. In this thesis we consider two problems: recognition of object categories (a car, a pedestrian) and recognition of object instances (Mr Smith's car, Mr Smith himself). We use local object representations, which means that an image is considered as a set of local regions, which is more robust and more flexible that a global representation. We particularly focus on bag-of-words methods, that discard geometric information between local regions. We study the influence of each step of the algorithm, and show that the parameter the most influent on the accuracy is the amount of local regions sampled to describe the image. We thus propose to sample a large amount of random local regions to describe images. In the context of this CIFRE industrial PhD thesis, in partnership with INRIA and Bertin Technologies, we study how performant bag-of-words methods are for recognizing military vehicles on infrared images. We show that the algorithm parameters have the same behavior as the ones in the visible spectrum. We also study operation parameters, such as the distance between the camera and the target, and show that the most critical parameters are the occlusion rate and the amount of textured background in the region of interest when targets are poorly segmented. We also study the trade-off between accuracy and computation time, and we propose a feature selection scheme well suited for multiclass hierarchical classifiers, more interesting than standard feature selection for flat classifiers. The three previous studies focus on object category recognition. We also consider object instance recognition, and we propose a similarity measure for comparing objects never seen during a training phase. That measure is based on the quantization by extremely randomized clustering forests of matching pairs of local regions sampled from the two images to compare. All these studies are validated by many experiments on state of the art and our own datasets, and we always obtain results as good as the state of the art, if not better.
Document type :
Complete list of metadatas

Cited literature [95 references]  Display  Hide  Download
Contributor : Eric Nowak <>
Submitted on : Thursday, July 24, 2008 - 3:08:08 PM
Last modification on : Thursday, November 19, 2020 - 1:00:20 PM
Long-term archiving on: : Saturday, November 26, 2016 - 12:35:05 AM


  • HAL Id : tel-00305664, version 1



Eric Nowak. Reconnaissance de catégories d'objets et d'instances d'objets à l'aide de représentations locales. Informatique [cs]. Institut National Polytechnique de Grenoble - INPG, 2008. Français. ⟨tel-00305664⟩



Record views


Files downloads