Publication detail

Brno University of Technology at TRECVid 2010

HRADIŠ, M. BERAN, V. ŘEZNÍČEK, I. HEROUT, A. BAŘINA, D. VLČEK, A. ZEMČÍK, P.

Original Title

Brno University of Technology at TRECVid 2010

Type

conference paper

Language

English

Original Abstract

This paper describes our approach to semantic indexing and content-based copy detection which was used for TRECVID 2010 evaluation. Semantic indexing 1.  The runs differ in the types of visual features used. All runs use several bag-of-word representations fed to separate linear SVMs and the SVMs were fused by logistic regression.    - F_A_Brno_resource_4: Only single best visual features (on the training set)      are used - dense image sampling with rgb-SIFT.    - F_A_Brno_basic_3: This run uses dense sampling and Harris-Laplace detector      in combination with SIFT and rgb-sift descriptors.    - F_A_Brno_color_2: This run extends F_A_Brno_basic_3 by adding dense sampling      with rg-SIFT, Opponent-SIFT, Hue-SIFT, HSV-SIFT, C-SIFT and opponent      histogram descriptors.    - F_A_Brno_spacetime_1: This run extends F_A_Brno_color_2 by adding space-time      visual features STIP and HESSTIP.      2. Combining multiple types of visual features improves results      significantly. F_A_Brno_color_2 achieve more than twice better results than      F_A_Brno_resource_4. The space-time visual features did not improve results.      3. Combining multiple types of visual features is important. Linear SVM is      inferior to non-linear SVM in the context of semantic indexing.      Content-based Copy Detection      1.    Two runs submitted, but with similar settings; the difference is only      in amount of processed test data (40% and 60%)         - brno.m.*.l3sl2: SURF, bag-of-words (visual codebook: 2k size, 4 nearest           neighbors used in soft-assignment), inverted file index, geometry           (homography) based image similarity metric      2.    What if any significant differences (in terms of what measures) did      you find among the runs?         - only one setting used - no differences      3.    Based on the results, can you estimate the relative contribution of      each component of your system/approach to its effectiveness?         - slow search in reference dataset due to unsuitable configuration of           used visual codebook      4.    Overall, what did you learn about runs/approaches and the research      question(s) that motivated them?         - change the way of describing the video content - frame based (or           key-frame based) approach is not sufficient

Keywords

TRECVID, semantic indexing, Content-based Copy Detection, image classification

Authors

HRADIŠ, M.; BERAN, V.; ŘEZNÍČEK, I.; HEROUT, A.; BAŘINA, D.; VLČEK, A.; ZEMČÍK, P.

Released

28. 12. 2010

Publisher

National Institute of Standards and Technology

Location

Gaithersburg, MD

Pages from

1

Pages to

10

Pages count

11

URL

http://www-nlpir.nist.gov/projects/tvpubs/tv10.papers/brno.pdf

BibTex

@inproceedings{BUT34908,
  author="Michal {Hradiš} and Vítězslav {Beran} and Ivo {Řezníček} and Adam {Herout} and David {Bařina} and Adam {Vlček} and Pavel {Zemčík}",
  title="Brno University of Technology at TRECVid 2010",
  booktitle="2010 TREC Video Retrieval Evaluation Notebook Papers",
  year="2010",
  pages="1--10",
  publisher="National Institute of Standards and Technology",
  address="Gaithersburg, MD",
  url="http://www-nlpir.nist.gov/projects/tvpubs/tv10.papers/brno.pdf"
}