CornerNet: Detecting Objects as Paired Keypoints

TL;DR

CornerNet introduces a novel object detection method that identifies objects as paired keypoints using a single CNN, eliminating anchor boxes and achieving state-of-the-art results on MS COCO.

Contribution

It proposes detecting objects as paired keypoints and introduces corner pooling, a new pooling layer, to improve corner localization.

Findings

Achieves 42.2% AP on MS COCO

Outperforms existing one-stage detectors

Introduces corner pooling for better corner detection

Abstract

We propose CornerNet, a new approach to object detection where we detect an object bounding box as a pair of keypoints, the top-left corner and the bottom-right corner, using a single convolution neural network. By detecting objects as paired keypoints, we eliminate the need for designing a set of anchor boxes commonly used in prior single-stage detectors. In addition to our novel formulation, we introduce corner pooling, a new type of pooling layer that helps the network better localize corners. Experiments show that CornerNet achieves a 42.2% AP on MS COCO, outperforming all existing one-stage detectors.