Discriminatively Activated Sparselets

TitleDiscriminatively Activated Sparselets
Publication TypeConference Paper
Year of Publication2013
AuthorsGirshick, R., Song H. Oh, & Darrell T.
Other Numbers3438
Abstract

Shared representations are highly appealingdue to their potential for gains in computationaland statistical efficiency. Compressinga shared representation leads to greatercomputational savings, but can also severelydecrease performance on a target task. Recently,sparselets (Song et al., 2012) were introducedas a new shared intermediate representationfor multiclass object detection withdeformable part models (Felzenszwalb et al.,2010a), showing significant speedup factors,but with a large decrease in task performance.In this paper we describe a new trainingframework that learns which sparselets toactivate in order to optimize a discriminativeobjective, leading to larger speedup factorswith no decrease in task performance. WeFirst reformulate sparselets in a general structuredoutput prediction framework, then analyzewhen sparselets lead to computationalefficiency gains, and lastly show experimentalresults on object detection and imageclassification tasks. Our experimental resultsdemonstrate that discriminative activationsubstantially outperforms the previousreconstructive approach which, togetherwith our structured output prediction formulation,make sparselets broadly applicableand significantly more effective.

URLhttps://www.icsi.berkeley.edu/pubs/vision/discriminativelyactivated13.pdf
Bibliographic Notes

Proceedings of the International Conference on Machine Learning (ICML 2013), Atlanta, Georgia

Abbreviated Authors

R. Girshick, H. O. Song, and T. Darrell

ICSI Research Group

Vision

ICSI Publication Type

Article in conference proceedings