Can I retrain a oneformer model only for semantic seg?

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

Closed RYHSmmc closed 1 year ago

praeclarumjj3 commented 1 year ago

Hi @RYHSmmc, thanks for your interest in our work. Yes, you can. There are two cases:

If you have panoptic annotations: Set TASK_PROB.SEMANTIC as 1.0 in the config. https://github.com/SHI-Labs/OneFormer/blob/56799ef9e02968af4c7793b30deabcbeec29ffc0/configs/ade20k/Base-ADE20K-UnifiedSegmentation.yaml#L56
If you only have semantic annotations: You can use the custom semantic dataset mapper.

praeclarumjj3 commented 1 year ago

Feel free to re-open if you are still facing any issues.