A completely affine invariant image-matching method based on perspective projection

In many cases, feature-matching problems can boil down to the computation of affine invariant local image features. However, many methods which are used to obtain these image features are based on affine not perspective trans-formation. They typically fail to get enough matching points at extreme viewpoints. In this paper, a novel method based on perspective projection to simulate all image views by sampling a camera pose in 3D space is presented. Only four variables-three angles (pan, tilt and roll) and a scale factor which are used to describe the camera pose - are involved to represent an affine transformation matrix. All these samples generated by our method are more similar to the real images captured by camera than those generated by traditional methods. We demonstrate the performance and robustness of our method by comparing it with popular SIFT, ASIFT descriptors and randomized trees method. Experimental results show that a large range of viewpoints by learning the behavior of key points patterns can be handled even when the camera is placed at some extreme positions.