how to get greyscale annotation images?

[PennTao]

The original ADE20K dataset does not provide greyscale annotation image, while the PSPNet needs greyscale annotation images for evaluation, is there are script that can convert original images o greyscale?

Or should I just add rgb2gray in eval_acc to turn all color imAnno into geryscale ?

btw, if I don't apply rgb2gray to imAnno, I will get matrix dimension mismatch error while doing IoU

[imjiangjun]

I met the same problem. This may be helpful:
http://groups.csail.mit.edu/vision/datasets/ADE20K/
You can download and run the tool to get the annotation images, which are index number for the colormap within png:
http://groups.csail.mit.edu/vision/datasets/ADE20K/code.zip

[qizhuli]

rgb2grey won't work because eval_acc does not need a grey image per se; it needs an indexed image, where an index corresponds to a class ID.

[navinranjan7]

PALETTE or Colormap doesn't match with the mask (annotation) while converting the 'RGB mask' to a single channel class ID. Has anyone encountered the same problem?
I tried multiple approaches to convert to a single-channel Class ID mask.

1. np.all(np.equal(RGB_mask, colormap[i]))
2. cv2.inRange(RGB_mask, colormap[i], colormap[i])

[navinranjan7]

http://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip
Can download the ADEChallengeData2016 dataset with annotations for 150 classes.
https://cv.gluon.ai/build/examples_datasets/ade20k.html