I am currently working on a project related to Multi label segmentation. I haven't been able to find any substantial papers where objects in images were segmented based on a membership function. For instance, if the model were to segment a car, the entire object of car would get segmented in the first place and then the various parts of the car would get segmented like its tires, mirrors etc. In this case, tires would have a membership-based value where it would belong in both the categories of car and the wheel.
To extract shapes such as the tires (which are mostly circles and ellipses), I am looking at the work done on Feature Extractors and Hierarchical Object Detectors in images. Most of the papers that I was able to find were published in early 2010s. Those papers made use of Deformable Parts Model, HOG and other filter or template matching based approaches. These methods don't utilize neural nets to learn these features and there haven't been any recent papers on these topics.
I was able to solve the problem of multi class segmentation using u-net but not in the method described above. For the multi class problem, each label belonged to single class whereas for multi label the problem is different. Does anyhow any experience working with multi label segmentation? Whenever I have searched for multi label segmentation always multi label classification has come up. Can anyone guide me with this?