Semantic Segmentation for objects with random shape

Hi All the examples and tutorial about Semantic Segmentation are about shape specific objects (e.g. cars, animals, scenes, plants, etc.). I am working on a problem in which the classes do not have a specific shape. They are visually distinguishable based on their textures. Now, my questions is whether semantic segmentation concept and methods (like CNN) can be used in such a context. Thanks, Roohollah

Answer 1

[Image Analyst]

I don't believe so. If there is no shape that your object takes on, and can be any random shape, then you'll have to use other methods to segment it, like texture segmentation. Then just don't worry at all about assigning a name to the shape.

Answer 2

[jim peyton]

The problem you're describing is texture segmentation. Here's a good tutorial on using Gabor filters to extract texture features (spatial frequency and orientation).

MathWorks: Texture segmentation with gabor filters

There are other types of texture filters you could use to extract features: Energy/Entropy, gradient magnitudes and directions, even creating your own shape-specific filters.