Deep Q-learning for Active Recognition of GERMS: Baseline performance on a standardized dataset for active learning

Mohsen Malmir, Karan Sikka, Deborah Forster, Javier Movellan and Garison Cottrell

Abstract

In this paper, we introduce GERMS, a dataset designed to accelerate progress on active object recognition in the context of human robot interaction. GERMS consists of a collection of videos taken from the point of view of a humanoid robot that receives objects from humans and actively examines them. GERMS provides methods to simulate, evaluate, and compare active object recognition approaches that close the loop between perception and action without the need to operate physical robots. We present a benchmark system for active object recognition based on deep Q-learning (DQL). The system learns to actively examine objects by minimizing overall classification error using standard back-propagation and Q-learning. DQL learns an efficient policy that achieves high levels of accuracy with short observation periods.

Session

Poster 2

Files

Extended Abstract (PDF, 1825K)
Paper (PDF, 6M)

DOI

10.5244/C.29.161
https://dx.doi.org/10.5244/C.29.161

Citation

Bibtex

@inproceedings{BMVC2015_161,
	title={Deep Q-learning for Active Recognition of GERMS: Baseline performance on a standardized dataset for active learning},
	author={Mohsen Malmir and Karan Sikka and Deborah Forster and Javier Movellan and Garison Cottrell},
	year={2015},
	month={September},
	pages={161.1-161.11},
	articleno={161},
	numpages={11},
	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
	publisher={BMVA Press},
	editor={Xianghua Xie, Mark W. Jones, and Gary K. L. Tam},
	doi={10.5244/C.29.161},
	isbn={1-901725-53-7},
	url={https://dx.doi.org/10.5244/C.29.161}
}