Anytime Recognition of Objects and Scenes

TitleAnytime Recognition of Objects and Scenes
Publication TypeConference Paper
Year of Publication2014
AuthorsKarayev, S., Fritz M., & Darrell T.
Other Numbers3710
Abstract

Humans are capable of perceiving a scene at a glance,and obtain deeper understanding with additional time. Similarly,visual recognition deployments should be robust tovarying computational budgets. Such situations requireAnytime recognition ability, which is rarely considered incomputer vision research. We present a method for learningdynamic policies to optimize Anytime performance invisual architectures. Our model sequentially orders featurecomputation and performs subsequent classification. Crucially,decisions are made at test time and depend on observeddata and intermediate results. We show the applicabilityof this system to standard problems in scene and objectrecognition. On suitable datasets, we can incorporatea semantic back-off strategy that gives maximally specificpredictions for a desired level of accuracy; this provides anew view on the time course of human visual perception.

Acknowledgment

This work was partially supported by funding provided to ICSI through National Science Foundation grant IIS : 0905647 (“Computer Vision and Online Communities: A Symbiosis”). Addition funding was provided through NSF grants IIS : 1134072 (``Support for Workshop on Advances in Language and Vision'') and IIS : 1212798 (``Reconstructive recognition: Uniting statistical scene understanding and physics-based visual reasoning''); through DARPA Mind’s Eye and MSEE programs; and by Toyota; and by the Intel Visual Computing Institute. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of the funders.

URLhttp://www.icsi.berkeley.edu/pubs/vision/anytimerecognition14.pdf
Bibliographic Notes

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Columbus, Ohio

Abbreviated Authors

S. Karayev, M. Fritz, and T. Darrell

ICSI Research Group

Vision

ICSI Publication Type

Article in conference proceedings