keyword

corner pooling

Corner pooling is a specialized pooling layer in convolutional neural networks designed to accurately detect and localize the corners of bounding boxes for object detection. Because the corner of an object frequently lies in the background outside the object itself and lacks distinct local visual cues, standard convolutional operations often struggle to recognize it. Corner pooling addresses this challenge by aggregating maximum feature values along orthogonal boundary directions, such as scanning horizontally to the right and vertically downward for a top-left corner, or scanning horizontally to the left and vertically upward for a bottom-right corner. By combining these directional maximal responses, the layer enables keypoint-based vision models to identify the spatial extent and boundaries of objects reliably without depending on predefined anchor boxes.

1 item

CornerNet: Detecting Objects as Paired Keypoints

Hei Law, Jia Deng

OrganizationsPrinceton University

Why you should read this

Proposes a keypoint-based object detection framework that eliminates the need for anchor boxes by predicting paired bounding box corners with specialized corner pooling, achieving top performance among single-stage detectors on MS COCO.

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.

Source

https://arxiv.org/pdf/1808.01244

Added

2026-09-11