CenterNet: Keypoint Triplets for Object Detection

TL;DR

CenterNet tackles the high rate of incorrect boxes in one-stage keypoint detectors by introducing a center keypoint triplet (center plus two corners) and two lightweight pooling modules that enrich internal object information. The model verifies corner pairs by the presence of a center keypoint within a scale-aware central region and optimizes both precision and recall with center pooling and cascade corner pooling. On MS-COCO, CenterNet achieves 47.0% AP on test-dev, outperforming all prior one-stage detectors and approaching two-stage detector performance, while maintaining efficient inference. The work offers a practical, plug-in enhancement to existing one-stage, anchor-free detectors to reduce false positives with minimal cost.

Abstract

In object detection, keypoint-based approaches often suffer a large number of incorrect object bounding boxes, arguably due to the lack of an additional look into the cropped regions. This paper presents an efficient solution which explores the visual patterns within each cropped region with minimal costs. We build our framework upon a representative one-stage keypoint-based detector named CornerNet. Our approach, named CenterNet, detects each object as a triplet, rather than a pair, of keypoints, which improves both precision and recall. Accordingly, we design two customized modules named cascade corner pooling and center pooling, which play the roles of enriching information collected by both top-left and bottom-right corners and providing more recognizable information at the central regions, respectively. On the MS-COCO dataset, CenterNet achieves an AP of 47.0%, which outperforms all existing one-stage detectors by at least 4.9%. Meanwhile, with a faster inference speed, CenterNet demonstrates quite comparable performance to the top-ranked two-stage detectors. Code is available at https://github.com/Duankaiwen/CenterNet.

Figures (7)

• Figure 1: In the first row, we visualize the top 100 bounding boxes (according to the MS-COCO dataset standard) of CornerNet. Ground-truth and predicted objects are marked in blue and red, respectively. In the second row, we show that correct predictions can be determined by checking the central parts.
• Figure 2: Architecture of CenterNet. A convolutional backbone network applies cascade corner pooling and center pooling to output two corner heatmaps and a center keypoint heatmap, respectively. Similar to CornerNet, a pair of detected corners and the similar embeddings are used to detect a potential bounding box. Then the detected center keypoints are used to determine the final bounding boxes.
• Figure 3: (a) The central region when $n=3$. (b) The central region when $n=5$. The solid rectangles denote the predicted bounding boxes and the shaded regions denote the central regions.
• Figure 4: (a) Center pooling takes the maximum values in both horizontal and vertical directions. (b) Corner pooling only takes the maximum values in boundary directions. (c) Cascade corner pooling takes the maximum values in both boundary directions and internal directions of objects.
• Figure 5: The structures of the center pooling module (a) and the cascade top corner pooling module (b). We achieve center pooling and the cascade corner pooling by combining the corner pooling at different directions.