CornerNet: Detecting Objects as Paired Keypoints

CornerNet proposes a new approach to object detection in which an object's bounding box is detected as a pair of keypoints, the top-left corner and the bottom-right corner, produced by a single convolutional neural network. By formulating detection this way, the method eliminates the need to design the sets of anchor boxes that are commonly used in prior single-stage detectors. To support this formulation, the authors introduce corner pooling, a new type of pooling layer that helps the network better localize corners.

Experiments show that CornerNet achieves 42.2% AP on the MS COCO benchmark, outperforming all existing one-stage detectors. The result demonstrated that an anchor-free, keypoint-based formulation could match and even surpass anchor-based single-stage detection, with corner pooling contributing to more accurate corner localization.