The initial scores for this paper were: 5: Marginally below the acceptance threshold. 4: An okay submission, but not good enough; a reject. 7: A good submission; an accept. The main critiques of the negative reviewers were lack of clarity and missing details. R2 had also doubts about novelty. The positive reviewer also pointed out to the many missing details but overall thinks the paper has interesting technical contributions validated by detailed experimental evaluation. The authors have provided the rebuttal. In the follow-up discussion the positive reviewer (R3) keeps their positive rating trusting the authors to provide some of the details still missing after the rebuttal in the final version of the work. R3 argues for accepting the work. R2 after reading the rebuttal and the discussion (see below) upgrades their score to 6: Marginally above the acceptance threshold. R1 in the discussion clarifies their position regarding novelty, thinks the paper is borderline, keeps their 5 rating, but is not against acceptance (see their comment below). The final scores for this work are: 5: Marginally below the acceptance threshold.  6: Marginally above the acceptance threshold. 7: A good submission; an accept. This is a borderline case. In the end AC is convinced by the positive arguments of R3 supported by R2 and recommends to accept this work. Input from R2 in the discussion after reading the rebuttal: “I updated my review according to the rebuttal. I think authors did indeed clarify a lot of doubts I had about results and their significance, and I think that the improvement over SOTA is pretty significant at this point. The architectures devised for the task seem to improve a lot, setting a new SOTA on a trending task. I think the community will be happy to discuss this contribution. I gladly welcome the added explanations in regarding the SAFA / SPE modules. “ Input from R1 in the discussion after reading the rebuttal: With the "somewhat incremental contribution" I was referring to the use of polar coordinate transform which as such is not new but the application to this particular problem is new (as far as I am aware). Also, the other key contribution (SAFA) was not very well explained. Nevertheless, I think that the improvements outlined in the author feedback will clarify this. I still think that this paper is a borderline, considering that NeurIPS is very selective in general and there are typically more good papers submitted than can be accepted. As there is no borderline rating available I chose rating 5 originally but I am not against acceptance.