Beyond Spatial Pyramids: Receptive Field Learning for Pooled Image Features

TitleBeyond Spatial Pyramids: Receptive Field Learning for Pooled Image Features
Publication TypeConference Paper
Year of Publication2012
AuthorsJia, Y., Huang C., & Darrell T.
Page(s)3370-3377
Other Numbers3284
Abstract

In this paper we examine the effect of receptive field designson classification accuracy in the commonly adoptedpipeline of image classification. While existing algorithmsusually use manually defined spatial regions for pooling, weshow that learning more adaptive receptive fields increasesperformance even with a significantly smaller codebook sizeat the coding layer. To learn the optimal pooling parameters,we adopt the idea of over-completeness by startingwith a large number of receptive field candidates, and traina classifier with structured sparsity to only use a sparse subsetof all the features. An efficient algorithm based on incrementalfeature selection and retraining is proposed for fastlearning. With this method, we achieve the best publishedperformance on the CIFAR-10 dataset, using a much lowerdimensional feature space than previous methods.

URLhttp://www.icsi.berkeley.edu/pubs/vision/beyondspatial12.pdf
Bibliographic Notes

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2012), Providence, Rhode Island, pp. 3370-3377

Abbreviated Authors

Y. Jia, C. Huang, and T. Darrell

ICSI Research Group

Vision

ICSI Publication Type

Article in conference proceedings