why training using original CSPN implementation and inference using accelerated CSPN

JUGGHM / PENet_ICRA2021

ICRA 2021 "Towards Precise and Efficient Image Guided Depth Completion"

MIT License

325 stars 46 forks source link

why training using original CSPN implementation and inference using accelerated CSPN #16

Open dongliangcao opened 3 years ago

dongliangcao commented 3 years ago

Hi, thanks for your great work! I am wondering why during the training, the original CSPN is used. Because as you mentioned in the paper, the new implemented one is much faster.

JUGGHM commented 3 years ago

Thanks for your interest! Actually the accelerated CSPN is originally designed for dilated CSPN as the sample and stitch steps are time consuming. And the introduced unfolding operation is time saving yet GPU consuming. So here is the answer: the original CSPN is more friendly for devices with smaller GPU memory.