Local shape patch based object detection

We present a novel object detection framework that uses the local shape patches features combining the interclass global features information. A supervised local model learning architecture is proposed: a novel interest point descriptor is proposed and applied to detect the local shape patches, the local shape patches are formed by chains of several connected contour segments. Then the object local contour parts model are learned from a small set of training images. In order to make the local contour patch model is invariant to rotation and translation, the local contour patch descriptor is represented by modified shape context. By adding the edge orientation information and interclass global information, the power of differentiating mismatches is increased, especially detecting the objects existing similar parts. Both experiments on image feature point matching and object detection comparing with other feature descriptors are carried out in order to validate the proposed method.