Learning-AI

ATSS: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

May 2020

tl;dr: The gap between anchor-based and anchor-free methods lie in the sampling strategy.

Overall impression

The paper founds out that the main difference between anchor-based methods (such as RetinaNet) and anchor-free methods (FCOS) mainly lies in the definition of positive examples and negative examples.

This paper draws much inspiration with FCOS and shall be read together.

Both ATSS and YOLOF deal with topk anchors. ATSS focuses on dynamically adjusting the threshold to balance the pos/neg anchors based on topk anchors. YOLOF focuses on having balanced pos/neg samples, by ignoring pos samples beyond topk.

Key ideas

Technical details

Notes