Objects as Points

2019

• Introduction
• Architecture
• My thinking

  Introduction

Why anchor free?

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(NMS).

This Paper model an object as a single point — the center point of its bounding box. The detector
uses keypoint estimation to find center points and regress to all other properties, such as offset,
size.

Architecture

• Splat all truth keypoints onto heatmap
  with Gaussian kernel.
• To recover the discretization error cause by the output stride, the paper additional predict a
  local offset
• Regress to the object size.

My thinking

• This detector predict center of objects directly without densely sample, so it’s efficient.
  While FCOS need to sample since the center point of a grid can been seen as a special anchor box.
• I guess that this detector cannot work well on dense object detection and object with strange shape.