Learning-Deep-Learning

FSAF: Feature Selective Anchor-Free Module for Single-Shot Object Detection

April 2019

tl;dr: Introduced anchor-free module on top of FPN, which can be used alone or with anchor-based method (such as RetinaNet). The paper also proposes to select feature layers automatically instead of manually assign target according to object size.

Overall impression

FSAF has two main contribution: FS (online/learned feature selection) training method and AF (anchor free) path. It would be of interest to see how would RetinaNet improve just based on the FS method vs the original heuristic/manual feature selection. Also, how about combining the feature maps (like in Panoptic FPN) so that there is no need to do feature selection?

Key ideas

Technical details

Notes

Anchor boxes are designed for discretizing the continuous space of all possible instance boxes into a finite number of boxes with predefined locations, scales and aspect ratios.