Can I use pretrained network csp-darknet53-coco?

Hello guys, I want to train YOLOv4 detector and as shown in example on MathWorks page they used pretrained network csp-darknet53-coco… but i have my own 1865 annotated imaged which contains 5 different classes which i specified… so can i use this network or do i have to make my own network using deep network designer in matlab and edit darknet53 network there? Next I want to train Faster R-CNN and SSD detector and as shown in examples on MathWorks page these detectors use pretrained ResNet50 network … question is the same - can i use this network or do i have to edit the resnet50 network in matlab deep network designer and use my edited network? Thanks for answers :)

Answer 1

[Neha]

Hi Adrain,

I understand that you want to know if pre-trained network can be used to create a detector with your own dataset. For the YOLOv4 detector, you can fine-tune the pretrained csp-darknet53-coco network using transfer learning. Similarly, the pre-trained ResNet-50 can be used to train Faster R-CNN and SSD Detector. This involves replacing the last few layers of the network with new layers that are tailored to your specific task or just replacing the last layer with a new output layer. This can be done using the Deep Network Designer app or by modifying the network architecture using the MATLAB code. Please refer to Transfer Learning Using Pretrained Network - MATLAB & Simulink (mathworks.com) for more information.

Hope this helps!

[Adrian Kleffler]

Hello, thanks for answer. But do i have to fine tune it? Or can i just use these networks? Will it be fine?

[Neha]

Hi Adrian,

You can just replace the last layer of the network with a new output layer to implement transfer learning.