Convolutional Randomized Binary Features for Keypoints Recognition

Abstract

Randomized binary features have been used to recognize keypoints. However, the existing methods apply their sampling operations on raw or blurred images. The great success of convolutional neural networks has proved that convolution operator is a powerful feature extractor. So, we try to combine binary feature extractor with convolutional feature extractor to generate convolutional randomized binary features. In our method, we first generate multi-layer convolutional feature maps for an image and apply the pre-generated sampling operators to sample on theses convolutional feature maps. Finally, all the sampling values are binary encoded into bytes-like feature vector. The basic sampling operators of traditional binary features have only two points, which are not suitable for multi-layer images. While, the basic sampling operator we used to observe multi-layer convolutional feature maps is RID (Randomized Intensity Difference) operator. The strategy that applying RID to convolutional feature maps can improve binary feature quality. Our methods are compared with state-of-art methods on several image datasets. The experiment results show the excellent performance of our method.