Laplacian margin distribution boosting for learning from sparsely labeled data

Tao Wang*, Xuming He, Chunhua Shen, Nick Barnes

*Corresponding author for this work

    Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

    Abstract

    Boosting algorithms attract much attention in computer vision and image processing because of their strong performance in a variety of applications. Recent progress on the theory of boosting algorithms suggests a close link between good generalization and the margin distrubtion of the classifier \wrt a dataset. In this paper, we propose a novel data-dependent margin distribution learning criterion for boosting, termed Laplacian MDBoost, which utilizes the intrinsic geometric structure of dataset. One key aspect of our method is that it can seamlessly incorporate unlabeled data by including a graph Laplacian regularizer. We derive a dual formulation of the learning problem that can be efficiently solved by column generation. Experiments on various datasets validate the effectiveness of the new graph Laplacian based learning criterion on both supervised and unsupervised learning settings. We also show that the performance of our algorithm outperforms the state-of-the-art semi-supervised learning algorithms on a variety of inductive inference tasks, including real world video segmentation.

    Original languageEnglish
    Title of host publicationProceedings - 2011 International Conference on Digital Image Computing
    Subtitle of host publicationTechniques and Applications, DICTA 2011
    Pages209-216
    Number of pages8
    DOIs
    Publication statusPublished - 2011
    Event2011 International Conference on Digital Image Computing: Techniques and Applications, DICTA 2011 - Noosa, QLD, Australia
    Duration: 6 Dec 20118 Dec 2011

    Publication series

    NameProceedings - 2011 International Conference on Digital Image Computing: Techniques and Applications, DICTA 2011

    Conference

    Conference2011 International Conference on Digital Image Computing: Techniques and Applications, DICTA 2011
    Country/TerritoryAustralia
    CityNoosa, QLD
    Period6/12/118/12/11

    Fingerprint

    Dive into the research topics of 'Laplacian margin distribution boosting for learning from sparsely labeled data'. Together they form a unique fingerprint.

    Cite this