failure on KITTI

[NorthSummer]

I can't reproduce the same result as the tensorflow version... could you please provide related works on KITTI, thanks

[glingli]

That's simple to change the model to suit for KITTI. You can't use his checkpoint which is trained on the shapenet with 1024 points as input, it is too much for KITTI which only have ~150 points. You should establish your dataset and train again. Also, you should change the parameters of the pcn net at pcn.py.
I trained for about 100k iterations with 2 hours, and here is the result:

This is much better isn't it.

[RhythmOfTheRain-Byte]

Excuse me， I'd like to ask how to train the completion for the car category on KITTI. Also, I want to know how the ground truth for the completion is obtained during the model training process.
Thank you very much！