Object Detection Using Geometrical Block Structures

We propose a novel method of object detection in unconstrained, clustered scenes. Our method strongly benefits from object representation using geometrical structure of image blocks. It comes from an intuition that object has strong relationships between some of its components. It effectively extends the features of local area to the global using a complete graph of blocks so as to achieve a perspective of features in geometrical structure of the object. AdaBoost is adopted to select those relations of block pairs which are able to distinguish object from the rest while designing classifier. This method gives a good result when we use face and human detection as testing cases.