DFR-net: Learning Dense Features at the Resolution Level

Patch matching is a significant stage in numerous computer vision tasks. This paper proposed a novel network structure, named DFR-net, appropriate to match patches. The proposed network uses a dense connectivity pattern at the resolution level, making the training efficient. This connectivity pattern has been shown improving the accuracy of patch matching. The DFR-net, with a single-tower architecture, focused on the relationship between (non-)corresponding patches, which improved the performance of the traditional Siamese network. The component of DFR-net, named RDCNet block, produces a smaller model size and is demonstrated suiting for patch matching. To ensure the experimental effectiveness, the DFR-net was trained on the public Brown patch dataset and the HPatches dataset.