Contour-Based Object Detection Using Max-Margin Hough Transform

In this paper, a contour-based object detection method based on Max-Margin Hough transform is proposed. We learn Implicit Shape Model using local contour features namely Pair of Adjacent Segments (PAS) features. A Max-Margin Hough transform (M²HT) [1] is then applied, where local parts generate weighted votes for possible object locations. Weights are learnt so that higher weights are assigned to parts which repeatedly appear in consistent locations. The achieved results on TUD cows reference dataset show that discriminative learning of weights improves the contour-based Hough detector.