coco stuff

[kevinwei0410]

hi, Im student in taiwan

Does the model not support for "coco-stuff" yet ?
the file : SenFormer\mmseg\core\evaluation\classesnames
which only have three different types of dataset,

dataset_aliases = {
'cityscapes': ['cityscapes'],
'ade': ['ade', 'ade20k'],
'voc': ['voc', 'pascal_voc', 'voc12', 'voc12aug']
}

...?

[WalBouss]

Hi,
The repository does support COCO-Stuff dataset. You can train SenFormer from scratch on COCO-Stuff or download pretrained weights from this repo and evaluate on COCO-stuff by running this command:

python tools/test.pysenformer_configs/senformer/coco-stuff10k/senformer_fpnt_r50_512x512_80k_coco.py /path/to/checkpoint_file

But you are right the class_names.py did not contain coco's classes and palette.
However, since it's only used in image_demo.py, it does not affect the rest of the repo.
Anyway, I have updated the affected files and image_demo.py should now work on COCO-Stuff too:

python demo/image_demo.py ${IMAGE_FILE} ${CONFIG_FILE} ${CHECKPOINT_FILE} --classes=coco --palette=coco

Hope that helps

[kevinwei0410]

hi,
Many thanks for your reply and the tutorials.

I run your model to do semantic segmentation on my custom data set, and the performance is impressive.
Therefore, I can get wonderful training data. It help me smoother to develop my final project.
In the final report, I will attach the link of your github and introduce how the training data was generated.

I hope everything goes well with your work and research.
best regards,

[WalBouss]

Glad it helped!

Good luck with your project, and feel free to reach out if you encounter additional complications with the code.
Best,