Scale up the input size

junyanz / interactive-deep-colorization

Deep learning software for colorizing black and white images with a few clicks.

https://richzhang.github.io/ideepcolor/

MIT License

2.69k stars 447 forks source link

Scale up the input size #62

Closed i-radwan closed 5 years ago

i-radwan commented 5 years ago

Hi there, Thanks for your great effort. I've a question related to the input shape, can we scale it up to colorize large input images (e.g. 1024 * 1024)? I guess the model won't fit in a 6 GB GPU, but can we eliminate some layers and still get good results (I can provide more hints to help the model colorize the large images)?

richzhang commented 5 years ago

A hack would be to process it in a sliding window fashion (colorize 256x256 blocks separately). This may give some discontinuities, however. You can also try to process overlapping windows. For example, first colorize the top-left img[0:256,0:256]. Then colorize img[192:448,0:256], by first seeding the overlapping region img[192:256,0:256] with some randomly selected colored points from the first colorization.