NeurIPS 2020

CoMIR: Contrastive Multimodal Image Representation for Registration

Meta Review

Overall, this paper received mixed reviews. The main argument against this paper (raised by both reviewer R1 and R3) is its novelty, as it currently seems to be primarily a variant of the infoNCE objective using data-augmentation that capture rotational invariances. After having read the reviews, author response and followed the reviewer discussion I would like to recommend acceptance. The main reason for this is that, as R 1 and R3 mention, this is a well-executed paper (perhaps better seen as an application paper), with good experiments and evaluation and also a method that tested and worked in 2 datasets, one of which in the biomedical domain.