search

cvlab-yonsei / dkn

An official implementation of "Deformable Kernel Network for Joint Image Filtering" (IJCV 2020) in PyTorch.
GNU General Public License v3.0
76 stars 9 forks source link

Downsampling method wrong !!! #5

Closed TZYSJTU closed 3 years ago

TZYSJTU commented 3 years ago

Well, after a month's study, I finnaly got the mistake point of the experiment.

It starts with the unmatched BIcubic upsampling baseline result: I found this because when I use Bicubic to downsample then use Bicubic to upsample resulting in much better RMSE 8.07 than 14.22 (8x upsampling) In the dataloader file, the authers use Bicubic to downsample the HR depth map to obtain LR input. target = np.array(Image.fromarray(depth).resize((w//s,h//s),Image.BICUBIC).resize((w, h), Image.BICUBIC))

However, all other papers use nearest neighbor downsampling on NYU v2. See CVPR paper "Pixel-Adaptive Convolutional Neural Networks" for evidence. This means the comparison is unfair. Therefore, all the experiment results should be re-trained with: target = np.array(Image.fromarray(depth).resize((w//s,h//s),Image.NEAREST).resize((w, h), Image.BICUBIC))

jun0kim commented 3 years ago

See the results of Table 3 in our paper.