I wonder if there is a method works as follows: 1) The algorithm is trained with some images of a particular object (I train my Image with 1000 banana images) 2) I use the training data to segment a test image (semantic segmentation)
The critical point here is, the training images DO NOT contain labeling information (foreground - background). I expect the algorithm to learn the shape of the object with some common edge orientations or features etc.
I study over Graphical Models for Segmentation and the most similar method is "OBJCUT" proposed by the following paper:
But, they train the system with shape exemplars (ground truth images contain segmentations done by humans).
My objects to segment are non-articulated objects.