optimal parameters for synthetic depth image to real depth image

[frk1993]

Hi, I am planning to train a cycleGAN model to translate synthetic depth images to real depth images from my depth-camera. I do not have any idea how to choose the optimal parameters. My synthetic and real images have size 480x640. I am planning to resize them to 512x512 -> load_size=512 and I will use crop_size=512. I have following questions: Is this a good idea to use cycleGAN for this problem ? Should I use a different crop_size? Should I change the network architecture? Should I change the default parameters for netD and netG?

Here example synthetic and real depth images from my dataset:

best regards

[junyanz]

Feel free to try CycleGAN.

• Crop_size: You can try 256 or 512.
• Use resnet_9blocks or resnet_6blocks.
• I will use the default ones.

[frk1993]

Thanks for your response. As you see if we compare synthetic image and real depth image the translation task is not just changing the color of the image there are also many small details (noise) in real image, which are missing in the synthetic image. And generally its shown that by increasing the number of resnet layers the performance can be increased. Therefore I am planning to extend the reset_9blocks to twelve. Would this be a good idea? And I have 500images from domain A and 500 hundred from domain B. How much epochs should I train ? Generally you used 200 epochs should I train for more epochs ?

Thanks and best regards

[junyanz]

reset_9blocks might be enough. But you can try 12. I used 200 epochs for most of the experiments. But you can try different epochs.