Motivation

• As we all know, strong Object Detection models so far all use the Anchor Box technique. 

• The most headache when using Anchor Box is the hyper-parameters to define the size and shape of the Anchor Box, whether the Anchor Box we choose is suitable for that Dataset? 

  1. Normally, these hyper-parameters will be selected through the k-means algorithm (from YOLOv2). However, the Anchor Boxes selected by those k-means are often already fitted on COCO, and sometimes won't work well with custom Datasets. 

  2. YOLOv5 cleverly handles this by using Genetic Algorithm (GA) after k-means to select an Anchor Box that better matches the custom Dataset. This adds a bit of time during training.

• With the use of Anchor Box, during training, we will have to perform IoU calculations to see which Anchor Box will correspond to which Ground Truth (GT) Box.

Objectives

• Label assignment is responsible for assigning labels to predefined anchors during the training stage. 

Related Works

• Simple IoU-based strategy and inside ground-truth method [41] 

• [5, 7, 18, 48, 51].

FCOS

• Different from anchor-based detectors, which consider the location on the input image as the center of (multiple) anchor boxes and regress the target bounding box with these anchor boxes as references, we directly regress the target bounding box at the location. 

• In other words, our detector directly views locations as training samples instead of anchor boxes in anchor-based detectors, which is the same as FCNs for semantic segmentation [20].

ATSS

• ATSS believes that the difference between Anchor-free and Anchor-based actually lies in how we define what is a positive sample (fore-ground/FG) and what is a negative sample (background/BG). And if positive samples and negative samples are well defined, then the accuracy of Anchor-free methods is comparable to Anchor-based methods.

• And this process of finding definitions for Positive and Negative samples is called the Label Assignment process.