Why the RPN proposal should add the ground truth?

[mutianwei521]

Dear Author,

I have read your wonderful research, containing paper and code. However, I found that, in the proposal box generated process, the boxes in RPN predicted are combined with the GT boxes. Why are you do that? If combine the GT box, I think the RPN will lost their function because I can directly use GT box to expand and train.

So, I have a question: why the RPN proposal do add the ground truth? If you can answer my question, I will appreciate.

Best wish

作者您好，

您这篇文章我读了之后，感受特别深刻。论文和代码都阅读了一下。但是我发现在代码中，您用了RPN预测的框和GT框做了合并，然后去进行expand和训练。如果说直接使用GT框，是不是不用RPN也是可以？

所以我的问题是：为啥要和GT框合并，意义在哪里？

如果您可以回答我的问题，我将不胜感激！