Why is the same color getting assigned to different regions of the an image?

[kanezaki]

Why is the same color getting assigned to different regions of the an image?

[zbyte64]

@kanezaki Speaking of adding edge information:
https://torchgeometry.readthedocs.io/en/latest/filters.html#kornia.filters.Sobel

[kanezaki]

No, the algorithm does not take into account "objectness" or any other kind of prior knowledge. By the way, if you want clusters of connected components, I found the following code useful.
https://github.com/davidstutz/superpixel-benchmark/blob/master/examples/cpp/connected_components.cpp

[pbamotra]

Also, in the example above can you help me figure out why the shoe was possibly labelled using two colors. It should ideally have one color. Also, looks like the SLIC is biased towards assigning same cluster to similarly colored objects in the image. Do you think that's the case?

[kanezaki]

Exactly. Since this method uses SLIC superpixels as the only clue for the image segmentation by CNN, there is no guarantee that two different superpixels are correctly merged. It would be interesting to add more clues (e.g., edges) to guide segmentation.

[pbamotra]

Yeah, I think if somehow connectivity can be added as a constraint to the objective then your method would be even better. =)
Thanks for prompt answers.