New RGB-D features for object recognition on kernel view

For the need of actually combining RGB data and depth input in computer vision research, new RGB-D features for object recognition are proposed. We present six kinds of RGB-D kernel matching functions on kernel view. They have the capability of capturing different RGB-depth cues including position, size, shape and distance. Due to the infinite dimensional character in Gaussian space, it is computationally expensive and unrealistic to utilize the kernel features directly. So we learn the compact RGB-D kernel descriptors using the methods of downsampling and PCA. The learned kernel features include: Spatio_dim kernel feature, Grad_nor kernel feature, Volum_shape kernel feature, and KPCA kernel feature. All of them can not replace each other, but complete each other. Extensive experimental results demonstrate that the proposed RGB-D kernel features are competitive, and that some of them outstrip other carefully crafted and sophisticated learned features, achieving 3-13% improvement in RGB-D recognition accuracy over the state of the art.