search

happinesslz / EPNet

EPNet: Enhancing Point Features with Image Semantics for 3D Object Detection(ECCV 2020)
MIT License
230 stars 37 forks source link

end-to-end training #5

Closed linbaiwpi closed 3 years ago

linbaiwpi commented 3 years ago

I noticed that during training, the train_mode is set to rcnn_online. It looks for me the authors choose train the rpn and rcnn networks end-to-end. Therefore, I am wondering why. Obviously, end-to-end training is easier but is it better than training in two steps?

happinesslz commented 3 years ago

@linbaiwpi Thanks for your attention. Yes, we train the network end-to-end. Maybe you can obtain the better model in two steps. However, it's hard for us to select the RPN model for better adapting the RCNN stage in two steps, which brings many parameters and a lot of uncertainty. So I am not recommended for training in two steps.