[Paper Review] CenterNet: Object Detection with Keypoint Triplets

Kaiwen Duan, Song Bai+4 more|arXiv (Cornell University)|Apr 17, 2019

Advanced Neural Network Applications35 references66 citations

TL;DR

CenterNet proposes a one-stage object detection framework that improves keypoint-based detection by modeling objects as keypoint triplets instead of pairs, using cascade corner pooling and center pooling to enhance feature representation. It achieves 47.0% AP on MS-COCO, outperforming all existing one-stage detectors while maintaining fast inference speeds comparable to top two-stage detectors.

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 extbf{47.0\%}, which outperforms all existing one-stage detectors by a large margin. 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.

Motivation & Objective

• To address the high rate of incorrect bounding boxes in keypoint-based object detectors by enhancing feature representation within cropped object regions.
• To improve both precision and recall in one-stage object detection without increasing computational cost significantly.
• To design a framework that leverages visual patterns in object regions more effectively than existing keypoint-based methods.
• To achieve performance on par with two-stage detectors while retaining the speed advantages of one-stage detectors.

Proposed method

• CenterNet models each object as a triplet of keypoints—specifically, a top-left corner, bottom-right corner, and a center keypoint—instead of a pair, to better localize objects.
• It introduces cascade corner pooling, which aggregates features from both top-left and bottom-right corners to enrich spatial context.
• It designs center pooling to enhance feature representation at the central region of objects, improving recognition of object centers.
• The method builds upon CornerNet as a base architecture, extending it with triplet-based detection and specialized pooling modules.
• The framework uses a single-stage detector with minimal additional computation, enabling high-speed inference.

Experimental results

Research questions

• RQ1 Can modeling objects as keypoint triplets instead of pairs improve detection accuracy in one-stage detectors?
• RQ2 How can visual patterns within cropped object regions be better exploited to reduce false positive bounding boxes?
• RQ3 Can specialized pooling modules like cascade corner pooling and center pooling enhance feature representation without increasing inference cost?
• RQ4 To what extent can a one-stage detector with triplet-based keypoint detection match or exceed the performance of two-stage detectors?

Key findings

• CenterNet achieves 47.0% AP on the MS-COCO dataset, setting a new state-of-the-art for one-stage object detectors.
• The proposed keypoint triplet approach improves both precision and recall compared to pair-based keypoint detectors like CornerNet.
• The addition of cascade corner pooling and center pooling significantly enhances feature representation in object regions.
• CenterNet maintains fast inference speed, demonstrating performance comparable to top-ranked two-stage detectors.