NeurIPS 2020

Distribution Matching for Crowd Counting

Meta Review

The initial ratings were 3577. The main concerns were: 1) contribution focused on a specific application of crowd counting; 2) missing comparisons; 3) missing ablation studies on hyperparameter selection. In the response, (1) authors argue they address a fundamental problem in spatial density estimation, which has broad impact, and they use crowd counting, which is is a well established CV task, for evaluation; (2) provide a table of comparisons showing improved results on large-scale datasets, NWPU and QNRF; (3) provide the ablation study results. After the response and discussion, R1 upgraded from 3 to 6, while R4 downgraded 7 to 6. R3's concerns were addressed and also upgraded to 6. The final ratings were 6667. After reading the reviews and responses, the AC agrees with the authors about the impact of the work on a fundamental problem in spatial density estimation, and also notes that reviewers' concerns were sufficiently addressed in the response. The paper offers new loss based on optimal transport, and provides a theoretical contribution of generalization bounds, and SOTA results. Thus the AC recommends acceptance. Authors should update the paper according to the reviews and responses.