Feature reduction for efficient object detection via L1-norm latent SVM

Abstract

The deformable part model is one of the most effective methods for object detection. However, it simultaneously computes the scores for a holistic filter and several part filters in a relatively high-dimensional feature space, which leads to low computational efficiency. This paper proposes an approach to select compact and effective features by learning a sparse deformable part model using L1-norm latent SVM. A stochastic truncated sub-gradient descent method is presented to solve the L1-norm latent SVM problem. Extensive experiments are conducted on the INRIA and PASCAL VOC 2007 datasets. Compared with the feature used in L2-norm latent SVM, a highly compact feature in our method can reach the state-of-the-art performance. The feature dimensionality is reduced to 12% in the INRIA dataset and less than 30% in most PASCAL VOC 2007 datasets. At the same time, the average precisions (AP) have almost no drop using the reduced feature. With our method, the speed of the detection score computation is faster than that of the L2-norm latent SVM method by 3 times. When the cascade strategy is applied, it can be further speeded up by about an order of magnitude. © Springer-Verlag 2013.