CenterNet: Keypoint Triplets for Object Detection This paper addresses the issue of incorrect bounding boxes in keypoint-based object detection, particularly in one-stage detectors like CornerNet. It introduces CenterNet, which detects objects using a triplet of keypoints (one center keypoint and two corner keypoints) instead of a pair, improving both precision and recall. CenterNet incorporates two custom modules: cascade corner pooling and center pooling, which enrich information from the top-left and bottom-right corners and the central region, respectively. On the MS-COCO dataset, CenterNet achieves an AP of 47.0%, outperforming all existing one-stage detectors by at least 4.9%. It also demonstrates comparable performance to top-ranked two-stage detectors with faster inference speeds. The code is available at https://github.com/Duankaiwen/CenterNet. Object detection has seen significant advancements with deep learning, particularly with one-stage approaches like CornerNet, which represent objects as pairs of corner keypoints. However, CornerNet struggles with global information and often generates incorrect bounding boxes, especially for small objects. CenterNet overcomes this by exploring visual patterns within each cropped region, using a triplet of keypoints. The approach includes center pooling to capture richer visual patterns and cascade corner pooling to enhance corner detection. These techniques improve the detection of center keypoints and corners, leading to more accurate bounding boxes. CenterNet is evaluated on the MS-COCO dataset, showing a significant reduction in false discovery rates and improved performance compared to CornerNet and other state-of-the-art detectors.CenterNet: Keypoint Triplets for Object Detection This paper addresses the issue of incorrect bounding boxes in keypoint-based object detection, particularly in one-stage detectors like CornerNet. It introduces CenterNet, which detects objects using a triplet of keypoints (one center keypoint and two corner keypoints) instead of a pair, improving both precision and recall. CenterNet incorporates two custom modules: cascade corner pooling and center pooling, which enrich information from the top-left and bottom-right corners and the central region, respectively. On the MS-COCO dataset, CenterNet achieves an AP of 47.0%, outperforming all existing one-stage detectors by at least 4.9%. It also demonstrates comparable performance to top-ranked two-stage detectors with faster inference speeds. The code is available at https://github.com/Duankaiwen/CenterNet. Object detection has seen significant advancements with deep learning, particularly with one-stage approaches like CornerNet, which represent objects as pairs of corner keypoints. However, CornerNet struggles with global information and often generates incorrect bounding boxes, especially for small objects. CenterNet overcomes this by exploring visual patterns within each cropped region, using a triplet of keypoints. The approach includes center pooling to capture richer visual patterns and cascade corner pooling to enhance corner detection. These techniques improve the detection of center keypoints and corners, leading to more accurate bounding boxes. CenterNet is evaluated on the MS-COCO dataset, showing a significant reduction in false discovery rates and improved performance compared to CornerNet and other state-of-the-art detectors.