Fully-Trainable Deep Matching
This work addresses the problem of making a popular image matching method fully trainable for researchers and practitioners in computer vision, though it is incremental as it builds on existing DM.
The paper tackled the limitation of Deep Matching not being trainable end-to-end by reformulating it as a convolutional neural network with a U-topology, resulting in improved performance on an image matching task.
Deep Matching (DM) is a popular high-quality method for quasi-dense image matching. Despite its name, however, the original DM formulation does not yield a deep neural network that can be trained end-to-end via backpropagation. In this paper, we remove this limitation by rewriting the complete DM algorithm as a convolutional neural network. This results in a novel deep architecture for image matching that involves a number of new layer types and that, similar to recent networks for image segmentation, has a U-topology. We demonstrate the utility of the approach by improving the performance of DM by learning it end-to-end on an image matching task.