A study on sampling strategies in space-time domain for recognition applications

Mert Dikmen, Dennis J. Lin, Andrey Del Pozo, Liang Liang Cao, Yun Fu, Thomas S Huang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

We investigate the relative strengths of existing space-time interest points in the context of action detection and recognition. The interest point operators evaluated are an extension of the Harris corner detector (Laptev et al. [1]), a space-time Gabor filter (Dollar et al. [2]), and randomized sampling on the motion boundaries. In the first level of experiments we study the low level attributes of interest points such as stability, repeatability and sparsity with respect to the sources of variations such as actors, viewpoint and action category. In the second level we measure the discriminative power of interest points by extracting generic region descriptors around the interest points (1. histogram of optical flow[3], 2. motion history images[4], 3. histograms of oriented gradients[3]). Then we build a simple action recognition scheme by constructing a dictionary of codewords and learning a recognition system using the histograms of these codewords. We demonstrate that although there may be merits due to the structural information contained in the interest point detections, ultimately getting as many data samples as possible, even with random sampling, is the decisive factor in the interpretation of space-time data.

Original languageEnglish (US)
Title of host publicationAdvances in Multimedia Modeling - 16th International Multimedia Modeling Conference, MMM 2010, Proceedings
Pages465-476
Number of pages12
DOIs
StatePublished - Dec 1 2009
Event16th International Multimedia Modeling Conference on Advances in Multimedia Modeling, MMM 2010 - Chongqing, China
Duration: Oct 6 2010Oct 8 2010

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume5916 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Other

Other16th International Multimedia Modeling Conference on Advances in Multimedia Modeling, MMM 2010
CountryChina
CityChongqing
Period10/6/1010/8/10

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'A study on sampling strategies in space-time domain for recognition applications'. Together they form a unique fingerprint.

Cite this