search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
541 stars 112 forks source link

About class weight in cityscapes. #22

Closed acgtyrant closed 6 years ago

acgtyrant commented 6 years ago

Is this right?

[ 1.5422179 6.2893033 10.492059 10.492059 10.492059 10.492059
10.492059 10.492059 10.492059 10.492059 10.492059 10.492059
10.492059 10.492059 10.492059 10.492059 10.492059 10.492059
10.492059 5.131662 ]

This means the first classes occurs 80% in labels, and almost the others class occurs less than 0.01%.

acgtyrant commented 6 years ago

OK, I did not convert the 255 label to 19 label, so the computed histogram is wrong.

Your program depend on these continuous labels, so the user must convert 255 to 19 so that they can run your program on Cityscapes successfully.