CenterNet: Keypoint Triplets for Object Detection

CenterNet tackles a known drawback of keypoint-based object detectors: they generate many incorrect bounding boxes because they lack an additional assessment inside the cropped regions. Built upon the representative one-stage keypoint-based detector CornerNet, CenterNet detects each object as a triplet of keypoints rather than a pair, which improves both precision and recall. To support this, the authors design two customized modules, cascade corner pooling and center pooling, that enrich the information collected by the top-left and bottom-right corners and provide more recognizable information from the central regions of objects.

On the MS-COCO dataset, CenterNet achieves an average precision (AP) of 47.0%, outperforming all existing one-stage detectors by at least 4.9%. Moreover, with a faster inference speed than the top-ranked two-stage detectors, CenterNet demonstrates performance comparable to those heavier detectors. This showed that adding a center-keypoint check to keypoint-based detection could substantially reduce false detections while keeping the efficiency advantages of one-stage approaches.