Unsupervised Modeling of Object Categories Using Link Analysis Techniques

People

Publication

  • Gunhee Kim, Christos Faloutsos and Martial Hebert
    Unsupervised Modeling of Object Categories Using Link Analysis Techniques
    IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2008), Alaska, USA, June 24-26, 2008. (Oral) (Acceptance = 62/1593~ 3.9%)
    [Paper(PDF)] [BibTeX] [Presentation(PPT)]

Matlab code

  • If you want to test my code (written in MATLAB), please email me.

Description

This paper proposes a novel unsupervised modeling of object categories in which (1) we first build a large-scale network that captures the interactions of all visual features across the image set, and (2) analyze the statistics of the link structure. Given a set of unlabelled images with a single number K, the number of object classes to be classified, our approach solves the following two subtasks: (1) Category discovery: clusters the input images into K groups according to object categories. (2) Localization: detects the most probable regions of the object in each classified image.

First our approach represents visual features in the form of a large-scale network (Fig.1.(a)). Then, the unsupervised classification and localization is formulated as the problem of finding hubs and communities in the network. The communities map to object categories and their hubs correspond to class-representative features. Two link analysis techniques are applied to mine those two types of information: (1) PageRank for ranking the relative importance of visual features, (2) Blondel et al's vertex similarity algorithm for structural similarity (Fig.1.(b)).

For category discovery, we observed that our approach achieved competitive performance against most of previous work. For localization, our approach fairly well discovered class representative features as hubs and suppressed the trivial information as background (See Fig.2).

representationstructural similarity
(a) A tiny part of the proposed network representation (b) Intution of structural similarity.
Figure 1. The large-scale network representation for unsupervised modeling.
results
Figure 2. Examples of localization. Our localization scheme behaves like feature selection. The first row shows the original extracted features (yellow elipses) in images and the second row represents the top 20% high-ranked features. The jet colormap is used to represent the importance weights from red (high) to blue (low).

Funding

  • Intelligent Robotics Development Program, a 21st Century Frontier R&D Programs by the Ministry of Commerce, Industry, and Energy of Korea.