Open sigmagod opened 5 years ago
What does "customize pre-trained result“ refer to?
Sorry, I follow the step from this repo and try to train a module from my datasets, its look like trainA and trainB to compression 256*256 So far, after trained complete how can I generate my customized pre-training(airplane2Eva.pb) to let the module transfer all the image what I want Or I need to follow this command below python test.py --dataroot datasets/horse2zebra/testA --name horse2zebra_pretrained --model test --name can change my dataset and names what I want? And I also follow the repo below, but the result is not very well, so I want to compare what the difference between https://github.com/vanhuyz/CycleGAN-TensorFlow By the way, for the dataset, the airplane position will be the key to determine the result? Or what should I notice?
Yes, you can use test.py
. Please see the instruction
It might be challenging to translate a CAD model to an entire image directly as it's hard to synthesize a background from a white background. You might need to segment the airplane and only run CycleGAN between segmented CAD model and segmented real images.
Hi, can you explain why a white background its hard to synthesize, so if I can synthesis CAD model below with some different picture, does it possibly will working?
CycleGAN assumes that images in different domains share similar content but with different styles. The white background has zero information and does not share content with the target domain.
@sigmagod Hi, how did you deal with this issue ? Did changing the backgrounds help you in the case ?
@sigmagod Hi, how did you deal with this issue ? Did changing the backgrounds help you in the case ?
@sudharavali Finally, I use style transfer and texturegan to rebuild CAD model texture, then synthesis with background.
Hi, Can I create my own customize pre-trained result for my train result dataset ?