Padding in Time-Consistency Loss

rozumden / DeFMO

[CVPR 2021] DeFMO: Deblurring and Shape Recovery of Fast Moving Objects

https://github.com/rozumden/DeFMO

MIT License

169 stars 24 forks source link

Padding in Time-Consistency Loss #3

Closed JLiu-Edinburgh closed 2 years ago

JLiu-Edinburgh commented 2 years ago

Hi,

Congratulations!

I found that "padding = tuple(side // 10 for side in sh[:2]) + (0,)" for normalized cross-correlation. Does it only implement padding to the height axis, since the padding tuple will be of size (4//10, H//10, 0)?

Thanks a lot.

rozumden commented 2 years ago

Hi,

Thanks for spotting this bug! I apologize for it, it appeared after some code cleaning for the camera-ready code release. It is fixed now.

Actually, in our recent improved and speeded-up reimplementation (https://github.com/klaegera-eth/DeFMO), we have the correct padding: https://github.com/klaegera-eth/DeFMO/blob/master/defmo/training/modules/loss.py (line 401).

Best regards, Denys