Jhonve / ImageSmoothing

Image smoothing via unsupervised learning
0 stars 0 forks source link

Flattening Criterion #1

Open luckyboy101 opened 4 years ago

luckyboy101 commented 4 years ago

flat_loss

output_copy = output.clone() flat_output = gaussian_kernel(output_copy) flat_map = (edge_map - 1.) -1. loss_flat = F.mse_loss(flat_map output, flat_map * flat_output) 这部分和原文的实现不一样,训练出来的效果不会好,对吧?

Jhonve commented 4 years ago

Yes. Actually, this repository is not a reproduction of paper 'Image Smoothing via Unsupervised Learning, Fan et al. ACM TOG (SIGGRAPH ASIA 2018)' (this repository is only my homework). Sorry for misleading and I will update the README.md later. 是的,和那篇siggraph asia的文章有区别,这里只是使用控制边界的方式做了一次尝试,效果确实没有原文的好。原文中的方法需要使用CUDA编写一个Operator,由于时间原因,没有来得及实现,后续有时间会继续更新。感谢您的关注!