A Maximum Margin Segmentation Selection for Visual Object Detection

Visual object detection is to predict the bounding box and the label of each object from the target classes in realistic scenes. Previous detection algorithms focus on training models to fit pre-segmented local patches. However, the patches themselves are not always meaningful due to the unsupervised segmentation mistakes. In this paper, a maximum margin method is proposed to get the optimal patches and the corresponding models simultaneously. The learning task is formulated as a quadratic programming (QP) problem and implemented in its dual form. When testing, we compute multiple segmentations of each image and select one segmentation with the maximum margin to predict their labels. We evaluate the detection performance of our algorithm on Pascal VOC2007 challenge data set and show some improved results with other detection algorithms.