Learning-Deep-Learning

CornerNet: Detecting Objects as Paired Keypoints

April 2019

tl;dr: Detect the top-left and bottom-right corner of the bbox, and learn an encoding for data association (associative embedding). It outperforms even multi-stage detectors such as mask rcnn and cascade rcnn.

Overall impression

The paper is the first anchor-less object detection paper in 2018 and rekindled people’s interest in anchor-less object detection framework. The corner pooling operation seems to have draw inspiration from the way humans draw bboxes. I would argue that the existence of corner pooling itself act as a proof that bbox is a bad representation for object detection.

This inspired later pioneering work such as CenterNet.

Key ideas

Technical details

Notes