FoveaBox: Beyond Anchor-based Object Detector

April 2019

tl;dr: RetinaNet based anchor-free method to predict center and top-left/bottom-right corners. It is more robust in detecting long objects.

Overall impression

This paper is one of the many recent papers about anchor-free object detection. FoveaBox and FSAF both build upon RetinaNet, but FSAF selects feature level by learning, but FoveaBox uses overlapping levels to detect an object. FoveaBox can also be used with anchor-based RetinaNet results, which performs better than FSAF.

Key ideas

Technical details