Why didn't fasterRCNNLayers implement alternating training?

Question

지현 2024-2-8

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2079756-why-didn-t-fasterrcnnlayers-implement-alternating-training

回答： Angelo Yeo 2024-2-17

I ran this script. (Object Detection Using Faster R-CNN Deep Learning)

I wanted to take a closer look at this layer(fasterRCNNLayers) by entering the following code[deepNetworkDesigner(lgraph)].

However, unlike what is in the paper(Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks), this layer did not implement alternating training.

Why is it not implemented?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Angelo Yeo 2024-2-17

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2079756-why-didn-t-fasterrcnnlayers-implement-alternating-training#answer_1410753

As far as I can understand, you probably have meant "alternating training" as the four steps in the paper, i.e.,

train the RPN, then freeze RPN layers,
train RCNN, then freeze RCNN layers,
train RPN, then freeze RPN layers
train RCNN.

It is implemented in Computer Vision Toolbox! You can do this by changing TrainingMethod to 'four-step' if you change the training method of trainFasterRCNNObjectDetector. See the document for detailed explanation for this option.