Locating Objects Without Bounding Boxes

December 2020

tl;dr: Locating objects with center points, but with a new loss function based on average Hausdorff distance.

Overall impression

This paper proposed a NMS-free method for object detection. The method by predicting a heat map is very similar to that of CenterNet and other anchor-free methods. However it proposes a new set-prediction loss based on the averaged Hausdorff distance.

Sometimes bbox of an object is not the optimal representation. Object localization is more appropriate where objects are very small, or substantially overlap. The bbox in crowded scenes may be infeasible to be used as GT.

Key ideas

Technical details