A Rapidly Trainable and Global Illumination Invariant Object Detection System

摘要

This paper addresses the main difficulty in adopting Viola-Jones-type object detection systems: their training time. Large training times are the result of having to repeatedly evaluate thousands of Haar-like features (HFs) in a database of object and clutter class images. The proposed object detector is fast to train mainly because of three reasons. Firstly, classifiers that exploit a clutter (non-object) model are used to build the object detector and, hence, they do not need to evaluate clutter images during training. Secondly, the redundant HFs are heuristically pre-eliminated from the feature pool to obtain a small set of independent features. Thirdly, classifiers that have fewer parameters to be optimized are used to build the object detector. As a result, they are faster to train than their traditional counterparts. Apart from faster training, an additional advantage of the proposed detector is that its output is invariant to global illumination changes. Our results indicate that if the object class does not exhibit substantial intra-class variation, then the proposed method can be used to build accurate and real-time object detectors whose training time is in the order of seconds. The quick training and testing speed of the proposed system makes it ideal for use in content-based image retrieval applications.