Separable Linear Classifiers for Online Learning in Appearance Based Object Detection

Online learning for object detection is an important requirement for many computer vision applications. In this paper, we present an iterative optimization algorithm that learns separable linear classifiers from a sample of positive and negative example images. We demonstrate that separability not only leads to rapid runtime behavior but enables very fast training. Experimental results underline that the approach even allows for real time online learning for tracking of articulated objects in real world environments. 1 Motivation and Scientific Context A general trend in present day computer vision research appears to be the integration of machine learning techniques into visual processing. Especially in the case of object detection in real world environments, the entanglement of vision and learning has led to stunning results. Cascaded weak classifiers rapidly detect objects of constraint shape and texture [1]. Taking aim at varying shape and texture, recent contributions simultaneously learn lexica of salient object parts as well as global structures [2–4]. Cognitive approaches integrate reasoning and learning across and within several levels of processing [5]. Robust as they are, the above techniques all require extensive training times. This hampers their use in scenarios where online learning is mandatory, as in the case of vision systems that assist their users in real world tasks. Among the few current proposals for such a scenario is a system that applies the Winnow algorithm for learning linear classifiers to motion data [6]. Others propose the use of sequential principal component analysis (PCA) and probabilistic tracking [7], or apply VPL classification, a technique that combines vector quantization, PCA and locally linear maps [8]. However, although they are fast, none of these methods reaches real time performance in online learning for object recognition. In this paper, we present a simple approach to very fast object learning which, nevertheless, provides rapid runtime behavior and reliable detection. Based on positive and negative example images, we propose an iterative least mean squares technique of learning separable linear classifiers. The method accomplishes input processing as rapidly as the popular cascaded weak classifiers. Moreover, it copes with objects of considerably varying shape and texture and is characterized by very short training times. Our classifiers therefore enable real time online learning in object recognition. The next section first discusses the benefits of linear classifiers for visual object detection and then introduces our algorithm for learning separable classifiers. Section 3 presents experimental results in online learning for object detection. Finally, a discussion ends this contribution. 2 Separable Linear Classifiers for Object Detection In their most basic form, binary linear classifiers compute the scalar product wx of a parameter vector w and a feature vector x. Their appeal for visual object detection lies in the fact that they may be implemented as two-dimensional linear filters. This requires writing parameters and features as matrices W and X and considering the Frobenius product of matrices W ⋆ X = ∑ i,j WijXij . If X denotes a digital image and W a suitable finite impulse response filter matrix of size m×n, a label yij characterizing the visual content in the vicinity of each pixel (i, j) can be computed from the convolution