ICS Action Database

Last update on 2009-03-24.
Update on 2007-12-01.
Update on 2006-11-01.
Update on 2004.4.1, 2.12,1.24,2003.9.17,6.22,6.6,3.12. Since 2003.3.10.

[Reference] Action Recognition Research of ICS

Japanese page

Usage of public data

These public data are intellectual properties of Intelligent Cooperative Systems Laboratory or Taketoshi MORI, Masamichi SHIMOSAKA and Kousuke TSUJIOKA. The public data are permitted to be used or changed, provided following the conditions below.

Before obtaining or using these public data, contact Moritmori@ics.t.u-tokyo.ac.jp. Especially, if you are planning to change the data, please contact including the fact.
If you use the data, write and specify cleary that you used "ICS Action Database" in the document, article or WWW page, and note URL http://www.ics.t.u-tokyo.ac.jp/. Also, present "Copyright (C) ICS, The University of Tokyo, 2003-2009".
Do not redistribute, duplicate distribute, sell, attach, exhibit or other similar act.
This public data set is provided as is without any guarantee or warranty. Take the risk by oneself.
Author and distribution manager do not share the responsibility for any damage or loss caused by usage or change of these public data. Take the risk by oneself.
Public data and the distribution method may be changed.
The intellectual property and permission condition are managed by Japanese law.
Author and distribution manager have right to change this permission condition.

Public data format

2004-01-24: Motion/Movement data are of BVH(BioVision Hiearhical) format, and action name data are added as evaluation value records.

2004.1.24 -- 2003.3.10 Public data:

The ground truth of action name for a motion is defined by "Human thought".
Action name may be usually assigned for a motion sequence. In spite of this, action name is assigned to every frame in the public data. A frame may correspond to several action names.

One dataset has 25 files.
The motion data are recorded in BVH(http://www.biovision.com/) format.
Action name data are recorded as evaluation values under tgb directory.
The action name is assigned based on questionnaire results. In tdb files, each frame has 1 for yes(The named action is occuring), 0 for no(The named action is not occuring), and 0.5 for neutral (Whether the named action is occuring or not is difficult to determine).

A frame of FoldArms file
Example of FoldArms file(AVI, About 300Kbytes)

A frame of GetUp file
Example of GetUp file(AVI, About 550Kbytes)

List of 25 files in a dataset

FoldArms	LookUp	Running	StandUp
GetUp	Lying	ShowingHand	Standing
Jump	LyingOnBack	SitDown	Turn
KeepDown	LyingOnFace	Sitting	Walking
LieDown	LyingOnSide	SittingOnChair
LookAway	OnFourLimbs	SittingOnFloor
LookDown	ReachOut	StandStill

There are 25 BVH files for target 25 actions.
Teacher data exist for target 25 actions for each BVH. i.e. 25 * 25 = 625 files.
There exist 5 above sets.

Public data

The data have been open to public since 2003-03-10.

2004-01-24 Public data

5 data: testData1 -- testData5. Each constructs a dataset.

Same as 2003-09-17 version. Compressed as zip.

2003-09-17 Public data

5 data: testData1～testData5. Each constructs a dataset.

[Reference] Action Recognition Research of ICS

Output example 1 of an ICS Action Recognition System

Related research papers

Kernel method and switching linear dynamics

Taketoshi Mori, Masamichi Shimosaka, Tatsuya Harada, Tomomasa Sato.
Time-Series Human Motion Analysis with Kernels Derived from Learned Switching Linear Dynamics
Transactions of the Japanese Society of Artificial Intelligence, Vol. 20, No. 3, 2005.
Masamichi Shimosaka, Taketoshi Mori, Tatsuya Harada, Tomomasa Sato.
Marginalized Bags of Vectors Kernels on Switching Linear Dynamics for Online Action Recognition
International Conference on Robotics and Automation, 2005.
Taketoshi Mori, Masamichi Shimosaka, Tatsuya Harada, Tomomasa Sato.
Time-Series Human Motion Analysis with Kernels Derived from Learned Switching Linear Dynamics.
The Second International Workshop on Man-Machine Symbiotic Systems(IWMMS2004),pp. 161-172, 10 2004.

Kernel discrimination and support vector learning

Taketoshi Mori, Masamichi Shimosaka, Tatsuya Harada, Tomomasa Sato.
Recognition of Daily Life Action and Its Performance Adjustment based on Support Vector Learning
International Journal of Humanoid Robotics, Vol.1, No.4, 565-583, 12 2004.
Taketoshi Mori, Masamichi SHIMOSAKA, Tatsuya Harada and Tomomasa Sato
Informative Motion Extractor for Action Recognition with Kernel Feature Alignment.
2004 IEEE/RSJ International Conference on Intelligent Robots and Systems,(IROS), pp.2009-2014, 9 2004.
Taketoshi Mori, Masamichi Shimosaka, Tomomasa Sato.
SVM-based Human Action Recognition and its Remarkable Motion Features Discovery Algorithm
The 9th International Symposium on Experimental Robotics 2004(ISER2004), 01paper129, 6 2004.

Kernel-based learning and qualitative prior knowledge

Masamichi Shimosaka and Taketoshi Mori and Tatsuya Harada and Tomomasa Sato.
Action Recognition Based on Kernel Machine Encoding Qualitative Prior Knowledge
Proceedings of 2004 IEEE International Conference on Systems, Man and Cybernetics, pp.1569-1576, The Hague, Netherlands, 10 2004.

Online action recognition and segmentation

Shimosaka, Nejigane, Mori, Sato.

Segmentation considering body structure

Taketoshi Mori, Yushi Segawa, Masamichi Shimosaka, Tomomasa Sato.
Segmentation of Sequential Daily-Actions Based on Hidden Markov Models and Human Body Hierarchy
The Second International Workshop on Man-Machine Symbiotic Systems(IWMMS2004), pp. 207-218, 10 2004.

Hidden Markov model and action hierarchy

Taketoshi Mori, Yushi Segawa, Masamichi Shimosaka, and Tomomasa Sato
Hierarchical Recognition of Daily Human Actions Based on Continuous Hidden Markov Models
The 6th International Conference on Automatic Face and Gesture Recognition(FG2004), 5 2004. to appear, accepted.

Japanese page

ICS Home Page

Copyright (c) ICS, the University of Tokyo, 2003-2009. wwwadmin@ics.t.u-tokyo.ac.jp
Copyright (c) Taketoshi Mori, 2003-2009. tmori@ics.t.u-tokyo.ac.jp