Using a custom dataset to train the Mobile-Seed Model

[yinheyanxian]

Hello, I would like to use my custom dataset to train your model.
My dataset is annotated using LabelMe, and I ultimately converted it into mask images suitable for training in MMsegmentation by referencing the Cityscapes dataset.
How can I obtain the semantic boundary ground truth required for training Mobile-Seed?
Thank you!

[martin-liao]

Thanks for your interest. You could refer to the label_generator here. We use this code to generate semantic edge maps for the CamVid and PASCAL datasets.
For each semantic class, we examine the local neighborhood of each pixel to check for the presence of pixels with different labels. Pixels whose neighborhoods contain heterogeneous labels are classified as semantic edge pixels for that class.