COCO-stuff for DetectoRS

[enemni]

\Hello,

when running python mmdetection/tools/misc/browse_dataset.py I got this error

mmdetection/mmdet/datasets/pipelines/loading.py", line 372, in _load_semantic_seg
    filename = osp.join(results['seg_prefix'],
  File "/usr/lib/python3.8/posixpath.py", line 76, in join
    a = os.fspath(a)
TypeError: expected str, bytes or os.PathLike object, not NoneType

In my config file I uncommented seg_prefix='data/coco/stuffthingmaps/train2017 since I was not able to generate a COCO-stuff version of my data. Is the COCO-stuff for the DetectoRS required? If yes as explained in the documentation, how can I create it for my own dataset?

thanks a lot,

[TaWeiYeh]

I figured out how to make COCO-stuff data. Take a look at the thing and stuff classes here.
The index of that class is the RGB values. However, remember to abstract all indexes by one.
For example, the class of person has index 1. The value of a person class is 0.

(Update) The output stuff and things image should be a 2D image ONLY. That means we are saving a 2D array to an image. Unfortunately, Matplotlib could not save a 2D image. I use PIL package instead.
Here is my code,
output_img = np.zeros((height, width), dtype=np.int32)
result = Image.fromarray(output_img.astype(np.uint8))
result.save(seg_filepath)