Active object detection

Object detection is the automatic determination of image locations at which instances of a predefined object class are present. Numerous methods for object detection exist (e.g., (Viola and Jones, 2001; Fergus et al., 2006)), most of which scan a part of the image at some stage of the object-detection process. Until now, this scanning is performed in a passive manner: local image samples extracted during scanning are not used to guide the scanning process. We mention two main object-detection approaches that employ passive scanning here. The window-sliding approach to object detection (e.g., (Viola and Jones, 2001)) employs passive scanning to check for object presence at all locations of an evenly spaced grid. This approach extracts a local sample at each grid point and classifies it either as an object or as a part of the background. The part-based approach to object detection (e.g., (Fergus et al., 2006)) employs passive scanning to determine interest points in an image. This approach calculates an interestvalue for local samples (such as entropy of gray-values at multiple scales (Kadir and Brady, 2001)) at all points of an evenly spaced grid. At the interest points, the approach extracts new local samples that are evaluated as belonging to the object or the background. Although some methods try to limit the region of the image in which passive scanning is applied (e.g., (Murphy et al., 2005)), it remains a computationally expensive and inefficient scanning method: at each sampling point computationally costly feature extraction is performed, while the probability of detecting an object or suitable interest point can be low. In this article, we investigate an object detection method that employs active scanning (based on (de Croon and Postma, 2006)). In active scanning local samples are used to guide the scanning process: at the current scanning position a local image sample is extracted and mapped to a shifting vector indicating the next scanning position. The method takes successive samples towards the expected object location, while skipping regions unlikely to contain the object. The goal of active scanning is to save computational effort, while retaining a good detection performance. In a companion article, we address the importance of our approach in the context of Embodied Cognitive Science (X). In this article we focus on the practical applicability in computer vision. In particular, we verify whether the method reaches its goal for a real-world task of face detection that has been studied before in (Kruppa et al., 2003; Cristinacce and Cootes, 2003). We compare the method’s performance and computational complexity with that