Why do we set a custom bias for classifier head in Object Detection?

Part 2 (2019)
1

I am studying Object detection and was going through the 2018 part 2 SSD notebook and the new Retinanet notebook. Both implementations set a custom bias for classifier head of object detection. In SSD, the bias is -3, and in Retinanet, it is -4.
SSD

ssd%20bias%20define
ssd bias define.png789×262 22.6 KB

ssd%20bias%20call
ssd bias call.png729×480 38.7 KB

Retinanet
retinanet%20bias%20define
retinanet bias define.png1015×244 30.6 KB

retinanet%20bias%20call

What is the reason for setting this custom bias for classifier head? Also, is there any specific reason behind the numbers -3 and -4?

2

Bias initialization comes from the Focal Loss paper

2 Likes