Multi-instance local exemplar comparisons for pedestrian detection

We propose to use the partial similarity between a sample and a number of exemplars as the image features for visual object detection. Define a part of the object as a sub-window inside the object bounding box, for each part of the object, a codebook of local appearance templates is learned. By using multiple templates for each part, and allowing the template to be compared with a bag of part instances in the neighborhood of the canonical location, the deformable and multi-aspect properties can be captured. A linear classifier is learned with feature selection, selecting a subset of the templates. To improve the efficiency of the detector, a rejection cascade is built by calibrating the linear classifier; the rejection cascade makes decisions using partial scores. Experimental results show that our method substantially improves the performance for human detection.