Dense geometric matching is a challenging computer vision task, requiring accurate correspondences under extreme variations in viewpoint and illumination, even for low-texture regions. In this task, finding accurate global correspondences is essential for later refinement stages. The current learning based paradigm is to perform global fixed-size correlation, followed by flattening and convolution to predict correspondences. In this work, we consider the problem from a different perspective and propose to formulate global correspondence estimation as a continuous probabilistic regression task using deep kernels, yielding a novel approach to learning dense correspondences. Our full approach, Deep Kernelized Matching, achieves significant improvements compared to the state-of-the-art on the competitive HPatches and YFCC100m benchmarks, and we dissect the gains of our contributions in a thorough ablation study.

No abstract

No abstract

No abstract

No abstract