Non-local methods exploiting the self-similarity of natural signals have been well studied, for example in image analysis and restoration. Existing approaches, however, rely on k-nearest neighbors (KNN) matching in a fixed feature space.

Update: I am somewhat convinced by the rebuttal. I will increase my rating, although I think that the quantitative improvements are quite marginal. Summary: Authors propose a neural network layer based on attention-like mechanism (a "non-local method") and apply it for the problem of image restoration. The main idea is to substitute "hard" k-NN selection with a continuous approximation: which is essentially a weighted average based on the pairwise distances between the predicted embeddings (very similar to mean-shift update rule). Although the paper is clearly written, the significance of the technical contributions is doubtful (see weaknesses), thus the overall score is marginally below the acceptance threshold.

Non-local methods exploiting the self-similarity of natural signals have been well studied, for example in image analysis and restoration. Existing approaches, however, rely on k-nearest neighbors (KNN) matching in a fixed feature space. To overcome this, we propose a continuous deterministic relaxation of KNN selection that maintains differentiability w.r.t. To exploit our relaxation, we propose the neural nearest neighbors block (N3 block), a novel non-local processing layer that leverages the principle of self-similarity and can be used as building block in modern neural network architectures. We show its effectiveness for the set reasoning task of correspondence classification as well as for image restoration, including image denoising and single image super-resolution, where we outperform strong convolutional neural network (CNN) baselines and recent non-local models that rely on KNN selection in hand-chosen features spaces.