Training a multi-exit cascade with linear asymmetric classification for efficient object detection

Abstract

Efficient visual object detection is of central interest in computer vision and pattern recognition due to its wide ranges of applications. Viola and Jones'detector has become a de facto framework [1]. In this work, we propose a new method to design a cascade of boosted classifiers for fast object detection, which combines linear asymmetric classification (LAC) into the recent multi-exit cascade structure. Therefore, the proposed method takes advantages of both LAC and the multi-exit cascade. Namely, (1) the multi-exit cascade structure collects all the scores of prior nodes for decision making at the current node, which reduces the loss of decision information; (2) LAC considers the asymmetric nature of the node training. We also show that the multi-exit cascade better meets the assumption of LAC learning than the standard Viola-Jones'cascade, both theoretically and empirically. Experiments confirm that our method outperforms existing methods such as Viola and Jones [1] and Wu et al. [2] on the MIT+CMU test data set.