1 Introduction

Anchor free object detection is powerful because of its speed and generalizability to other computer vision tasks.

“CenterNet: Objects as Points” is one of the milestones in the anchor-free object detection algorithm.

Anchor free object detection is more generalizable for different tasks: top for object detection, middle for 3D object detection, bottom: for pose estimation.

2 Method

1. The model has three output heads: the keypoint heatmap 128x128x80, object size 128x128x2, and local offset 128x128x2.
   1. The depth of the keypoint heatmap is C=80 (number of object classes)
2. Let us assume a point p(px,py) is the center of a bounding box. The output stride of the CenterNet model is R=4, so the point in the heatmap space will be p~:

2.1 GT heatmap & Loss

splat all ground truth keypoints onto a heatmap Y

(Left) Ground Truths of different classes(Represented in different colors). (Right) Three centers of respective classes splat into heatmaps using Gaussian Kernel.

a penalty-reduced pixelwise logistic regression with focal loss [33]: