SHI-Labs / OneFormer

OneFormer: One Transformer to Rule Universal Image Segmentation, arxiv 2022 / CVPR 2023
https://praeclarumjj3.github.io/oneformer
MIT License
1.39k stars 129 forks source link

About the problem of training with heterogeneous data. #84

Closed hhaAndroid closed 10 months ago

hhaAndroid commented 11 months ago

Hello, author. First of all, thank you very much for open-sourcing such excellent code. OneFormer can jointly train multiple segmentation tasks, but the dataset source is the same. I think it is difficult for any data to have panoramic, instance, and semantic segmentation annotations at the same time. Is it possible to extend it to data from multiple sources, but each source only includes annotations for specific tasks? I don't know if you have considered or tried this aspect.

praeclarumjj3 commented 11 months ago

Hi @hhaAndroid, thanks for your interest in our work. You can also train OneFormer on single-task annotation by writing your custom dataset mappers. We also provide some sample custom dataset mappers for users' convenience. https://github.com/SHI-Labs/OneFormer/tree/main/datasets/custom_datasets

praeclarumjj3 commented 10 months ago

I am closing this issue due to inactivity. Feel free to re-open.