Category-level detection based on object structures

We present a new class of descriptors which exhibit the ability to yield meaningful structural description of the object. These descriptors are constructed by harnessing the geometrical relationships and spatial configurations between two types of image primitives: Quadrangles and ellipses. Specifically, we extract the line segments from the line edge map of the image and exploit the spatial qualities of the line segments and the salient colors of the image to construct the quadrangles. The ellipses are extracted with a close loop system that is driven by Gestalt Psychology. Experimental results show very good performance for category-level object detection in which the objects in each category exhibit variations in form, scale and viewpoint.