Reviews: Neighbourhood Consensus Networks

Summary: This paper propose a differentiable and an end-to-end approach to refine the correspondences (both semantic and instance level) by utilizing the neighborhood consensus on 4D correlation matrix (of feature map). The pipeline is evaluated for PASCAL keypoints dataset for semantic correspondences, and InLoc for instance level correspondences Pros: intuitive, well described approach (I have some implementation level questions though, please see below) clearly state of the art on previously used benchmarks. Cons: - Computational Time is linear with the increase in resolution. How to handle the scenarios where dense correspondences between a pair of HD images (1080x1920) is required? How much time does it take to compute dense correspondences between two images by the proposed approach?