Detail publikace
Brno University of Technology at TRECVid 2008
CHMELAŘ, P. BERAN, V. HEROUT, A. HRADIŠ, M. JURÁNEK, R. LÁNÍK, A. MLÍCH, J. NAVRÁTIL, J. ŘEZNÍČEK, I. ŽÁK, P. ZEMČÍK, P.
Originální název
Brno University of Technology at TRECVid 2008
Anglický název
Brno University of Technology at TRECVid 2008
Jazyk
en
Originální abstrakt
In this paper we describe our experiments in all task of TRECVid 2008. This year, we have concentrated mainly on the local (affine covariant) image features and its transformation into a search-able form for the Content-based copy detection pilot together with the indexing and search techniques for the Search task and a practical test of the background subtraction and trajectory generation algorithms for the Surveillance pilot. In brief, we have submitted the following tasks: 1. Surveillance event detection pilot. We have participated in the detection of the following events - PersonRuns, ObjectPut, ElevatorNoEntry and OpposingFlow. It has been based mainly on advanced masking and background subtractions and extracted trajectories. 2. Content-based copy detection pilot. We have submitted one run based on search of the joint image features - global (color, texture) and local features (SIFT). 3. High-level feature extraction. We have used two training methods based on SVM using color, texture and face image features. First only selected subset of the training data, second all the annotated data were used for the training. 4. Search. We have performed two fully automatic IR experiments based on the text of the queries and ASR/MT provided by NIST and the data consumed by the High-level feature extraction task. 5. Rushes summarization, to which is dedicated a separate paper [2].
Anglický abstrakt
In this paper we describe our experiments in all task of TRECVid 2008. This year, we have concentrated mainly on the local (affine covariant) image features and its transformation into a search-able form for the Content-based copy detection pilot together with the indexing and search techniques for the Search task and a practical test of the background subtraction and trajectory generation algorithms for the Surveillance pilot. In brief, we have submitted the following tasks: 1. Surveillance event detection pilot. We have participated in the detection of the following events - PersonRuns, ObjectPut, ElevatorNoEntry and OpposingFlow. It has been based mainly on advanced masking and background subtractions and extracted trajectories. 2. Content-based copy detection pilot. We have submitted one run based on search of the joint image features - global (color, texture) and local features (SIFT). 3. High-level feature extraction. We have used two training methods based on SVM using color, texture and face image features. First only selected subset of the training data, second all the annotated data were used for the training. 4. Search. We have performed two fully automatic IR experiments based on the text of the queries and ASR/MT provided by NIST and the data consumed by the High-level feature extraction task. 5. Rushes summarization, to which is dedicated a separate paper [2].
Dokumenty
BibTex
@inproceedings{BUT29657,
author="Petr {Chmelař} and Vítězslav {Beran} and Adam {Herout} and Michal {Hradiš} and Roman {Juránek} and Aleš {Láník} and Jozef {Mlích} and Jan {Navrátil} and Ivo {Řezníček} and Pavel {Žák} and Pavel {Zemčík}",
title="Brno University of Technology at TRECVid 2008",
annote="In this paper we describe our experiments in all task of TRECVid 2008. This year,
we have concentrated
mainly on the local (affine covariant) image features and its transformation into
a search-able form for the
Content-based copy detection pilot together with the indexing and search
techniques for the Search task and a
practical test of the background subtraction and trajectory generation algorithms
for the Surveillance pilot.
In brief, we have submitted the following tasks:
1. Surveillance event detection pilot. We have participated in the detection
of the following events
- PersonRuns, ObjectPut, ElevatorNoEntry and OpposingFlow. It has been
based mainly on
advanced masking and background subtractions and extracted
trajectories.
2. Content-based copy detection pilot. We have submitted one run based on
search of the joint
image features - global (color, texture) and local features (SIFT).
3. High-level feature extraction. We have used two training methods based on
SVM using color,
texture and face image features. First only selected subset of the
training data, second all the
annotated data were used for the training.
4. Search. We have performed two fully automatic IR experiments based on the
text of the queries
and ASR/MT provided by NIST and the data consumed by the High-level
feature extraction task.
5. Rushes summarization, to which is dedicated a separate paper [2].",
address="National Institute of Standards and Technology",
booktitle="Proceedings of TRECVID 2008",
chapter="29657",
edition="NEUVEDEN",
howpublished="print",
institution="National Institute of Standards and Technology",
year="2008",
month="november",
pages="1--16",
publisher="National Institute of Standards and Technology",
type="conference paper"
}