xiaochus/MobileNetV3

Implementing Semantic Segmentation Head?

fzyzcjy opened this issue · 9 comments

Firstly, thanks for the implementation of the MobileNetV3!

Do you have an interest in implementing the semantic segmentation head described in section 6.4? I am currently planning to implement that, so if you have an interest, we can work together to do it!

@fzyzcjy Great idea, i'd like to work with you, but i don't have time in workday. You can fork my repository and write your code firstly, i will do it in my free time.

@xiaochus OK I will try that!

How do you suggest me on testing? My naive idea: 1. Train the classification (i.e. your existing code) on ImageNet 2. Implement segmentation, and train on Cityscapes, with weights initialized from ImageNet
Is that feasible, and what is your idea? In addition, do you have pretrained weights, or should I train by myself and upload that?

@fzyzcjy I don't have a pre-train model in imagenet. You can test the performance of mobilenetV3 in imagenet and it will be helpful in segmentation task.

OK I will make some attempt tonight

@fzyzcjy I add the code of Lite R-ASPP Semantic Segmentation head and you can check if i have a correct understanding of the description in the paper.

@xiaochus Oh I also did that since yesterday! But my code is currently still a big mess... Maybe I just post a gist