About training the model as mmentioned in your paper

[SamGhK]

Hello, sincere thanks for your great work and making the code public. I have a question regarding the training of the model. In your paper, you have mentioned:

We first train the 2D detector, along with the backbone, for 120K iterations using the Adam
optimizer. Then the 3D reasoning modules, IDE, 3D localization and local corners, are trained > for 80K iterations with the Adam optimizer. Finally, we use SGD to optimize the whole network
in an end-to-end fashion for 40K iterations.

Does this means that training for 3d BBoxes Detection should be as follow:
In FastBox/Optimizer set refine = True for 120k iteration then set joint_3d, depth, location and corners = True for 80k iterations. Finally set joint_2d_3d=True with Optimizer 'SGD' for 40k iterations ?

Could you please give me a feedback regarding the training progression of the model? How should the steps be configured?
I appreciate your input in this regard and look forward to hearing from you.

[Zengyi-Qin]

Thank you for your interest!

I would suggest to train as follows:

1. joint_2d_3d = True, others = False, for 80K iterations, Adam.
2. joint_2d_3d = True, others = False, for 40K iterations, SGD.

[Zengyi-Qin]

In the paper we wrote our original training procedural. But the abovementioned procedural is simpler and can also work

[SamGhK]

Thanks for your response. I will check the procedure you have mentioned.