search

ZM-Zhou / SDFA-Net_pytorch

Apache License 2.0
17 stars 5 forks source link

image cropping #10

Closed ksh11023 closed 1 year ago

ksh11023 commented 1 year ago

Hello,

Does cropping the image to (192x640) produces better result compared to directly using (384x1280) for training?

Thank you.

ZM-Zhou commented 1 year ago

Hi ksh11023, There are two main advantages of using image cropping in my opinion: 1. saving the GPU memory; 2. the discrete depth prediction could benefit from the random resizing and croping augmentations (mentiond in FAL-Net). I didn't try to use the raw high-resolution image for training due to GPU limitations, so I'm not sure which strategy is better.

ksh11023 commented 1 year ago

Thank you for the reply!