search

Dootmaan / DSRL

Implementation of CVPR 2020 Dual Super-Resolution Learning for Semantic Segmentation
MIT License
98 stars 17 forks source link

deeplab.py line 80 #3

Open L1Uc opened 4 years ago

L1Uc commented 4 years ago

为什么返回时将x_sr_up重复了两次呢？

L1Uc commented 4 years ago

根据train.py里调用的信息来看应该是self.point_wise(x_sr_up)吧？

Dootmaan commented 4 years ago

Hi there, thank you for your question. x_sr_up is returned twice only to meet the [segmentation_output, super-res_output, segmentation_FA_input, super-res_FA_input] format. And currently, super-res_output is just the same as super-res_FA_input so you will see the second and fourth position of return are both x_sr_up. I have also tried using self.pointwise(x_sr_up) as the FA input but then the result is a mess. Actually, in the paper the authors only mentioned that they used a feature transform module (and that is self.pointwise()) on SSSR path and didn't say much about the SISR path, so by the results of my experiments I don't think we need feature transform module for x_sr_up.