Discriminative Hough forests for object detection

Abstract

Object detection models based on the Implicit Shape Model (ISM) use small, local parts that vote for object centers in images. Since these parts vote completely independently from each other, this often leads to false-positive detections due to random constellations of parts. Thus, we introduce a verification step, which considers the activations of all voting elements that contribute to a detection. The levels of activation of each voting element in the ISM form a new description vector for an object hypothesis, which can be examined in order to discriminate between correct and incorrect detections. We learn classifiers to discriminate correct and wrong part constellations and thus assign a better confidence to each detection using linear models as well as a histogram intersection kernel. Additionally, we show how to use the discriminative weights of the linear classifier, not only as a post processing step, but directly in the voting process. We evaluate the proposed approach on different datasets, showing considerable improvements of the detection results.