Multi-Path Marginal Space Learning for Object Detection

Abstract

This chapter introduces a novel method for fast detection of objects with a large number of parameters. The method is based on Marginal Space Learning (MSL), which is a learning-based optimization technique that approaches the search for objects in images as a particle filter in a chain of subspaces of increasing dimensions, using trained detectors to prune the particles in the subspaces. MSL has been used extensively in Medical Imaging for detecting organs and landmarks in 2D and 3D data and for detecting and tracking curve-like structures such as guidewires and catethers. This chapter brings three contributions. Firstly, it introduces multiple computational paths for MSL, which can improve the detection performance compared to a single MSL path. Second, it presents an application of multi-path MSL to four parameter face detection from grayscale images. Thirdly, it observes experimentally that multiple-path MSL obtains a compact classifier with good generalization abilities. Consequently, the number of training examples can be reduced to half compared to other methods with similar performance.