CenterNet++ for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection is crucial for computer vision tasks.
• Current state-of-the-art methods predominantly use top-down approaches.
• Bottom-up approaches offer potential for improved recall rates.

Purpose of the Study:

• To demonstrate the competitive performance of bottom-up object detection.
• To introduce CenterNet, a novel bottom-up object detection framework.
• To achieve state-of-the-art results using an anchor-free, keypoint-based method.

Main Methods:

• CenterNet detects objects as a triplet of keypoints: top-left corner, bottom-right corner, and center.
• Objects are localized by grouping corner keypoints and confirming with center keypoints.
• The approach is anchor-free, eliminating the need for predefined anchor boxes.
• Adapted to various backbone architectures like 'hourglass' and 'pyramid' networks.

Main Results:

• CenterNet achieves state-of-the-art performance on the MS-COCO dataset, with APs of 53.7% (Res2Net-101) and 57.1% (Swin-Transformer).
• Outperforms all existing bottom-up object detectors.
• A real-time CenterNet model achieves 43.6% AP at 30.5 FPS, balancing accuracy and speed.

Conclusions:

• Bottom-up object detection, exemplified by CenterNet, is a viable and high-performing alternative to top-down methods.
• CenterNet's keypoint-based, anchor-free design enables robust object detection across scales and shapes.
• The framework offers a strong balance between detection accuracy and computational efficiency.