NeurIPS 2020

Detection as Regression: Certified Object Detection with Median Smoothing

Meta Review

The work proposes a simple and effective approach to certify object detection as a regression problem. The theoretical proofs that show the weakness of mean smoothing and effectiveness of the method are interesting. The approach also has potential to be used in other domains and applications. The author response was critical in the decision to accept this paper and next time I would recommend the authors evaluate on the full test set before paper submission. Overall the work would be a good contribution to the conference assuming the full test set evaluation, AP metrics and a discussion on inference speed are included.