search

GeorgeSeif / Semantic-Segmentation-Suite

Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic Segmentation models easily!
2.51k stars 878 forks source link

About label_image #111

Open Leo551 opened 6 years ago

Leo551 commented 6 years ago

Thanks for your great work. But when I try to train my own data sets, Do I need to convert the Unit-8 grey scale label images into color label images first?

GeorgeSeif commented 6 years ago

Yes you do! Although I'm thinking to create a script to do that for you since many people have requested. I'll post that soon!

Leo551 commented 6 years ago

Thanks!

zhangyunming commented 6 years ago

have you give the script?i want to train your code with CHIP dataset

x7hkvip commented 5 years ago

I convert the label image by myself,but when i training my own data,the ground truth in my ckpt files all class are same color(equals my csv files first line),what should i do?

burhr2 commented 5 years ago

I convert the label image by myself,but when i training my own data,the ground truth in my ckpt files all class are same color(equals my csv files first line),what should i do?

Hi, you can try to check with your color dict if they real match. maybe use .getcolor( ) and see your label color. Another thing, i have also tried on my own custom dataset but i do not get meaningful results, i m mostly get random segmentation output(with DeepLabv3, AdapNet and many others). As suggested by #168 this maybe caused by powerful models with small dataset(i.e need more data). I am curious to hear your result if they are meaningful