I'm wondering if there are techniques to extract features from a set of 2D binary masks and use them as a priori rules to constraint and refine an incoming mask.
To better understand what I'm talking about, let's call each mask of the set A_X (A_1, A_2...), and let's consider they are coming from the same 3D objects (slight variation in size and aspect ratios) with different planar projections. And let's call the incoming mask M, also coming from the same type of 3D object with a planar projection that has been already covered with the A_X set.
I've investigated the contours of the shape and the relationship between points of the contours to encode the "path" it follows so that I can refine the contours of the M but it didn't work well. I've also investigated the morphological techniques (erosion, dilation, smoothing, ...) but they tend to worse the incoming mask M because they is no rules to apply the morphological operation only to specific parts of the M.
I've already investigate the Statistical Appearance models but it seems that they require the extraction of landmarks from the incoming mask M to refine it, but my problem is that I'm not able to extract the landmarks from M as it is too noisy. (I think I can retrieve landmarks from A_X masks but I can't find any matching in M). Here is the link of a library (called Menpo) I've investigated.
I've also tried to perform registration (using ANTs for example) from the images in the "atlas" to the new image but here again if I only apply affine transformation, the obtained registered masks are too far from M, and when I apply deformation, I'm getting close to M but I only want rigid transformation as I know that the mask can't be deformed in my use case. This technique is a also computationally expensive as it requires several registration, but we can imagine ordering A_X based on the projection in a kind of decision-tree shape so that we can parse the tree based on the distance with each node (we perform a registration at each node of the tree and compare the distance with M and based on the distance we go left or right or we stop)
I'm going to investigate the Conditional Random Fields technique (as in this link) but the main problem is that I'm worried the raw image can't be used because post-processing has been applied to M and it doesn't follow the regular gradient of the raw image.
PS: There is no code in this question because I'm looking for techniques and not a particular way of solving a specific problem.