Cityscapes class weightings do not correspond to my calculations

sacmehta / EdgeNets

This repository contains the source code of our work on designing efficient CNNs for computer vision

MIT License

412 stars 82 forks source link

Cityscapes class weightings do not correspond to my calculations #17

Closed mitalbert closed 5 years ago

mitalbert commented 5 years ago

Hi,

Below are the class weights of Cityscapes train dataset I have calculated, which are different from ones found in train_segmentation.py. Can you please share the formula/script for calculating the class weights?

0.0819 0.4754 0.1323 1.5207 1.5190 2.4737 8.1900 5.2310 0.1869 1.4687 0.6881 1.9825 7.8143 0.4165 1.3811 1.1989 0.6276 5.3559 4.0925

sacmehta commented 5 years ago

We used it from ERFNet. See here:

https://github.com/Chenfeng1271/erfnet_pytorch/blob/09efaac1dc7829e3719552cbe1e63183368f916d/train/main.py#L113