guanfuchen / semseg

常用的语义分割架构结构综述以及代码复现 华为媒体研究院 图文Caption、OCR识别、图视文多模态理解与生成相关方向工作或实习欢迎咨询 15757172165 https://guanfuchen.github.io/media/hw_zhaopin_20220724_tiny.jpg
770 stars 164 forks source link

Label Refinement Network for Coarse-to-Fine Semantic Segmentation #2

Open guanfuchen opened 6 years ago

guanfuchen commented 6 years ago

预计实现LRN网络,未找到相关代码,进行相应的试验测试,架构和结构如下所示:

整体架构 RE精修模块

Surflyan commented 5 years ago

你好,请问LRN网络目前在复现吗,预计什么时候代码会开放出来,谢谢🙏。

guanfuchen commented 5 years ago

@Surflyan 最近在忙着写论文。。。刚实现了一下,你参考一下lrn.py,不过loss没有用多个loss,只是简单把最后的输出了,你把代码改一下自己复现试试,不知道效果怎么样,直接训练貌似可以,等训练几个周期应该就好了。

image

Surflyan commented 5 years ago

感谢,我找时间改一下。

guanfuchen commented 5 years ago

recurrence process

image for view image

loss epoch image

acc epoch image

mIoU on val epoch

image

guanfuchen commented 5 years ago

multi-loss mIoU is worse, just 0.58. the reason maybe is the resize for label and origin image, and use too much coarse loss.

image

image

image

image