Skip to main content
QUICK REVIEW

[Paper Review] Objects as Points

Xingyi Zhou, Dequan Wang|arXiv (Cornell University)|Apr 16, 2019
Advanced Neural Network ApplicationsComputer Science59 references739 citations
TL;DR

This paper proposes CenterNet, a novel object detection framework that models objects as single center points rather than axis-aligned bounding boxes. By using keypoint estimation to predict center points and regressing to object properties like size, 3D location, and orientation, CenterNet achieves state-of-the-art speed-accuracy trade-offs, attaining 28.1% AP at 142 FPS and 45.1% AP with multi-scale testing at 1.4 FPS on MS COCO.

ABSTRACT

Detection identifies objects as axis-aligned boxes in an image. Most successful object detectors enumerate a nearly exhaustive list of potential object locations and classify each. This is wasteful, inefficient, and requires additional post-processing. In this paper, we take a different approach. We model an object as a single point --- the center point of its bounding box. Our detector uses keypoint estimation to find center points and regresses to all other object properties, such as size, 3D location, orientation, and even pose. Our center point based approach, CenterNet, is end-to-end differentiable, simpler, faster, and more accurate than corresponding bounding box based detectors. CenterNet achieves the best speed-accuracy trade-off on the MS COCO dataset, with 28.1% AP at 142 FPS, 37.4% AP at 52 FPS, and 45.1% AP with multi-scale testing at 1.4 FPS. We use the same approach to estimate 3D bounding box in the KITTI benchmark and human pose on the COCO keypoint dataset. Our method performs competitively with sophisticated multi-stage methods and runs in real-time.

Motivation & Objective

  • To address the inefficiency and complexity of traditional object detectors that enumerate numerous candidate bounding boxes.
  • To simplify object detection by modeling each object as a single point—the center of its bounding box—reducing computational overhead.
  • To improve inference speed and accuracy by eliminating post-processing and enabling end-to-end differentiable training.
  • To extend the center-point paradigm to 3D detection and human pose estimation, demonstrating broad applicability.

Proposed method

  • The detector uses keypoint estimation to predict the center point of each object in the image.
  • For each detected center point, the network regresses to object properties including width, height, 3D location, orientation, and keypoint coordinates.
  • The method is end-to-end differentiable, allowing joint optimization of center point prediction and regression heads.
  • A single-stage architecture eliminates the need for region proposal networks or non-maximum suppression.
  • The model leverages feature pyramid networks and hourglass-style backbones for multi-scale feature learning.
  • The approach is adapted to 3D detection on KITTI and keypoint estimation on COCO by modifying the regression heads accordingly.

Experimental results

Research questions

  • RQ1Can object detection be effectively reformulated as a keypoint estimation problem centered on object centroids?
  • RQ2Does modeling objects as single points lead to faster inference and improved accuracy compared to traditional box-based detectors?
  • RQ3Can the center-point paradigm be generalized to 3D object detection and human pose estimation tasks?
  • RQ4How does the proposed method compare in speed-accuracy trade-off to existing state-of-the-art detectors on MS COCO and KITTI?

Key findings

  • CenterNet achieves 28.1% AP at 142 FPS on MS COCO, demonstrating a strong speed-accuracy trade-off.
  • At 52 FPS, the model attains 37.4% AP, outperforming many existing detectors at similar inference speeds.
  • With multi-scale testing, CenterNet reaches 45.1% AP on MS COCO, approaching the performance of more complex models.
  • The method performs competitively on 3D detection in the KITTI benchmark, achieving real-time inference.
  • CenterNet achieves state-of-the-art results on human keypoint estimation in the COCO dataset while maintaining real-time inference.
  • The end-to-end differentiable design eliminates the need for post-processing like non-maximum suppression.

Read this paper

Open the full text in the Nubint reader and ask about any passage as you read.

Read this paper on Nubint

This review was generated by AI from the original paper. Please check the paper before citing it.