SegNet Training w Custom Image Set Problems

[sarobando]

Hello,

I am trying to train SegNet with a custom set of images I labelled using CVAT. I followed the community tutorial linked in this thread. I am trying to run through it but encountering errors. Currently, I am unable to get the system to recognize my images in the split_custom.py step. I am using a set of PNG color images and the PNG mask images as inputs. I haven’t quite exhausted my options but I wanted to make sure that the guide above is still relevant before I sink too much time into it.

I recognize that this is perhaps outside the scope of this forum, but any advice would be appreciated. Additionally, I wanted to check if there might be a new accepted way to go about training on a custom dataset.

Thank you!

[dusty_nv]

Hi @sarobando, I’m not personally familiar, but it would appear that the split_custom.py script looks for jpg images by default, but you can change the extension to png by specifying the --image-format=png option.

Also I think that tool expects the dataset to have similar directory structure to Pascal VOC, so you may need to modify it it yours isn’t. Or to be honest, you may be able to skip splitting the datasets into train/val subsets if you just wanna get it up and running. Getting this working will probably require some hacking around with the scripts.