Introduction

Object detection is to detection the objects in an image. Generally, the type of objects are limited and pre-known. The task requires to output the bounding box (usually a rectangle box) of all objects.
The task is well-studied and own strong practical value in productions.

CNN based object detecion mainly consists of two stage network and one stage network. Current days, it seems the one stage network overwhelm the scenario.

One stage detector

SSD

Yolo

Retina-net(Focal loss)

FCOS

EfficientDet

ScratchDet

How to improve the detection network

1. ScratchDet The downsample in the Resnet stem is too high? Try 3x3 conv
   Discussion

2. FPN is used for multiple scale fusion.
   Try: [NAS-FPN] or [EfficientDet] or [PANet]
   refer: Blog

3. Loss function:
   try L1 norm, GIOU