CornerNet: Detecting Objects as Paired Keypoints

ECCV 2018  ·  Hei Law, Jia Deng ·

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.

PDF Abstract ECCV 2018 PDF ECCV 2018 Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Object Detection COCO minival CornerNet511 (Hourglass-104) box AP 38.4 # 182
AP50 53.8 # 102
AP75 40.9 # 86
APS 18.6 # 78
APM 40.5 # 75
APL 51.8 # 69
Object Detection COCO test-dev CornerNet511 (Hourglass-52, single-scale) box mAP 37.8 # 209
AP50 53.7 # 154
AP75 40.1 # 148
APS 17.0 # 139
APM 39.0 # 140
APL 50.5 # 137
Hardware Burden None # 1
Operations per network pass None # 1
Object Detection COCO test-dev CornerNet511 (Hourglass-104, multi-scale) box mAP 42.1 # 172
AP50 57.8 # 148
AP75 45.3 # 119
APS 20.8 # 131
APM 44.8 # 112
APL 56.7 # 83
Hardware Burden None # 1
Operations per network pass None # 1

Methods