Closed pmwenzel closed 5 years ago
Hi,
Each model always can do both directions, since that is integral to how it's trained.
But trust me, you don't want to go from day to night. It's not pretty...
I assumed the same since this is the logic CycleGANs are being trained. However, why do you think the way from day to night is so much worse than from night to day?
The night images from RobotCar are plagued with speckle nosie, compression-y artifacts, as well as this weird blue bounceback noise from the car's Lidar. Thus fake night images generated are super artifacty and ugly.
Do you also provide a pretrained model for robotcar_2night, which translates from day to night images? Thanks