Beyond HOG: Learning Local Parts for Object Detection

Abstract

Histogram of Oriented Gradients (HOG) features have laid solid foundation for object detection in recent years for its both accuracy and speed. However, the expressivity of HOG is limited because the simple gradient features may ignore some important local information about objects and HOG is actually data-independent. In this paper, we propose to replace HOG by a parts-based representation, Histogram of Local Parts (HLP), for object detection under sliding window framework. HLP can capture richer and larger local patterns of objects and are more expressive than HOG. Specifically, we adopt Sparse Nonnegative Matrix Factorization to learn an over-complete parts-based dictionary from data. Then we can obtain HLP representation for a local patch by aggregating the Local Parts coefficients of pixels in this patch. Like DPM, we can train a supervised model with HLP given the latent positions of roots and parts of objects. Extensive experiments on INRIA and PASCAL datasets verify the superiority of HLP to state-of-the-art HOG-based methods for object detection, which shows that HLP is more effective than HOG.