Issue reproducing test results

[erdenegb]

Hi I'm having an issue reproducing the table for U-Net I've downloaded the pretrained models but it seems like my mean and median angular error is consistently higher than what was reported in the paper, for the mixed case I have the following:

python main.py --mode 'test' --data_root '../galaxy_512' --camera 'galaxy' --checkpoint 'pretrained_galaxy/best.pt' --image_pool 1 2 3
MAE_illum : 2.8571817367076875 2.0290048122406006 12.943260192871094

python main.py --mode 'test' --data_root '../sony_512' --camera 'sony' --checkpoint 'pretrained_sony/best.pt' --image_pool 1 2 3
MAE_illum : 2.8872369016079524 2.6680153608322144 14.594013214111328

python main.py --mode 'test' --data_root '../nikon_512' --camera 'nikon' --checkpoint 'pretrained_nikon/best.pt' --image_pool 1 2 3
MAE_illum : 2.6096585162130057 2.207797050476074 8.505157470703125

This seems to be also true for just the single and multi illuminant test images. Am I missing something? Are the pretrained models different from the paper?