search

PRBonn / bonnet

Bonnet: An Open-Source Training and Deployment Framework for Semantic Segmentation in Robotics.
GNU General Public License v3.0
323 stars 89 forks source link

How to get datasets #64

Closed lccatala closed 4 years ago

lccatala commented 4 years ago

Hi, I'm trying to run this but can't figure out how to add the required datasets. Already tried it with the Cityscapes one, but it has a different structure than the one the system seems to require. Am I supposed to download them from somewhere and put them in a particular folder in a specific way?

tano297 commented 4 years ago

https://github.com/PRBonn/bonnet/blob/master/train_py/dataset/aux_scripts/cityscapes_preprocess.py

you have to use that script

tano297 commented 4 years ago

I suggest using this repo, as it is more up to date though... https://github.com/PRBonn/bonnetal

lccatala commented 4 years ago

Thanks! (I'm using this repo since it's the one my professors recommended, might try the new one afterwards). I've also found it seems the script https://github.com/PRBonn/bonnet/blob/master/train_py/dataset/aux_scripts/cityscapes_preprocess.py is broken (line 213), since it picks up every image ending in "_gtFine_labelTrainIds.png" as a label, and there are not such files in the dataset, thus giving me an empty /lbl/ folder. the only tails present are "_gtFine_labelIds.png" and "*_gtFine_instanceIds.png". Should I change it to one of those? And in that case, to which one?

tano297 commented 4 years ago

Before you can use the script in bonnet you need to create these labelids.png images using the scripts from the cityscapes API. You need to map the classes yourself using the dictionary in this file to make them match my yaml file

lccatala commented 4 years ago

Should the images only contain two classes? After preprocessing them they only have pixels for road and sidewalk, with most of them untagged (and it drops all of them)