search

JUGGHM / PENet_ICRA2021

ICRA 2021 "Towards Precise and Efficient Image Guided Depth Completion"
MIT License
326 stars 46 forks source link

Confusion about crop size #27

Open rightchose opened 3 years ago

rightchose commented 3 years ago

0000000000

rightchose commented 3 years ago

The image size is 1216x352? So follow your setting, the image feed in the model with size 1216 x352, but the top of the image is valid. The image I download from kitti depth completion rank.

rightchose commented 3 years ago

So I wonder why not use 1216x256?

rightchose commented 3 years ago

In my understanding, in main.py the args.val_h and args.val_w mean the size of area we crop from the image using bottom crop, and args.he and args.w is the size crop from the area.

JUGGHM commented 3 years ago

Hi! There are actually two settings. Ours follows the one of Ma et al.(Sparse2Dense) to feed the full sized images into the models. And the other one is cropping the images on the top into 1216x256 ones (e.g. GuideNet). Under this scheme, the top of the images are manually fulfilled when generating results on validation/test sets. Both of them are feasible while there might be some minor difference in performance.

rightchose commented 3 years ago

Oh! Thanks a lot. So comparing your model with others on the kitti website, your result may be different on the top of image.