Understanding and Improving Kernel Local Descriptors
This work addresses the need for robust and efficient local descriptors in computer vision, offering an incremental improvement over existing methods without requiring labeled data.
The authors tackled the problem of improving local patch descriptors by proposing a multiple-kernel descriptor based on pixel gradients, which combines polar and Cartesian parametrizations for robustness to mis-registration and uses whitening to enhance performance. Their unsupervised variant achieved competitive results against deep learning methods on various tasks.
We propose a multiple-kernel local-patch descriptor based on efficient match kernels from pixel gradients. It combines two parametrizations of gradient position and direction, each parametrization provides robustness to a different type of patch mis-registration: polar parametrization for noise in the patch dominant orientation detection, Cartesian for imprecise location of the feature point. Combined with whitening of the descriptor space, that is learned with or without supervision, the performance is significantly improved. We analyze the effect of the whitening on patch similarity and demonstrate its semantic meaning. Our unsupervised variant is the best performing descriptor constructed without the need of labeled data. Despite the simplicity of the proposed descriptor, it competes well with deep learning approaches on a number of different tasks.