About training with width-3 about photosketch HOT 2 CLOSED

TL;DR: you can try to thin the output in post-processing instead of re-training the model.

Hi, thanks for your interest. The width of the sketch is a tricky problem. First of all, training with width-3 is feasible, but might require a different lambda_A. I suggest you train first with the default width-5 for sanity check. Second, we use width-5 to train the model is because width-5 provides stronger supervision signal and thus better performance. If you want to get the same performance as width-5, you might want to consider sample balancing techniques used in boundary detection, e.g. a weight the loss by the inverse of the ratio between foreground and background pixels. Coming back to your task of getting thinner output, I would suggest you to apply some thinning operations in the post-processing step instead of re-training the model, for example, check out bwmorph.

from photosketch.

I get it! Thanks for your reply!

from photosketch.