L0-regularized Object Representation for Visual Tracking

View Researcher's Other Codes

Disclaimer: The provided code links for this paper are external links. Science Nest has no responsibility for the accuracy, legality or content of these links. Also, by downloading this code(s), you agree to comply with the terms of use as set out by the author(s) of the code(s).

Authors Jin-shan Pan, Jongwoo Lim, Zhixun Su, Ming-Hsuan Yang
Journal/Conference Name BMVC
Paper Category
Paper Abstract Introduction & Motivation: Visual tracking is a highly researched topic in the computer vision community since it has been widely applied in visual surveillance, driver assistant system, and many others. Although much progress has been made in the past decades, designing a practical visual tracking system is still a challenging problem due to numerous challenges in real world. Very recent efforts have been made to improve this method in terms of both speed and accuracy by using APG algorithm [1] or modeling the similarity between different candidates [6]. The works in [4, 5] point out that the aforementioned methods do not exploit rich and redundant image properties which can be captured compactly with subspace representations. Thus, they propose combining the strength of subspace learning [3] and sparse representation for modeling object appearance. In their work the raw pixel templates used in in [1, 2] are replaced with the orthogonal basis vectors (e.g., PCA basis), and the coefficients for an image are obtained by a least square (LS) method. However, we empirically find that such linear combination of the orthogonal basis vectors sometimes include redundant parts (e.g., background portions), which will interfere with the accuracy of object representation. We in this paper address this problem by proposing a tracking method based on an L0 regularized object representation. The L0 regularized object representation is able to reduce the redundant features while keeping the most important part. Furthermore, the estimation of the L0 regularized parameters can be efficiently conducted by the proposed APG algorithm.
Date of publication 2014
Code Programming Language MATLAB
Comment

Copyright Researcher 2021