How Yolo 3 is implemented in Yolo 4?

Question

I try to understand the architecture of Yolo4. It is composed of a backbone, neck, dense prediction and sparse prediction. Knowing that Yolo 3 has already a backbone, Is Yolo 4 taking all the architecture of Yolo 3 including its backbone or just part of Yolo3 ?

In page number 5 in paper Yolo 4, they've mentioned anchor based for Yolo3

Yolo 4 : https://arxiv.org/pdf/2004.10934.pdf

Yolo 3 :https://pjreddie.com/media/files/papers/YOLOv3.pdf

Answer 1

Yolov4 Vs Yolov3:

• Yolov3 uses Darknet53 as backbone, Yolov4 uses CSPDarknet53 as backbone.
• Yolov4 uses PANet as the method of parameter aggregation from different backbone levels for different detector levels, instead of the FPN used in Yolov3.

YOLOv4 consists of:

1. Backbone: CSPDarknet53 (Feature Extraction)
2. Neck: Additional module - SPP, PANet [this was not there in Yolov3]
3. Head: YOLOv3 (Dense Prediction Block) [This part is same as Yolov3]

Neck: Object detectors developed in recent years often insert some layers between backbone and head, and these layers are usually used to collect feature maps from different stages.

References:

• Read Yolov4 paper in detail
• Read this article