Today I talked with Bai about my plans for the further run. She suggests me that A (learning) method may be obtained through multi frames of the flow, predict point moving, and hence add a term in the energy function of matching. It’s a convinience thing that can be done directly from the ground true set, and coding could be started soon.
Also a person in this area is introduced: JiangXiong Yang, someone be doing stereo matching and flow for years. Plenty papers together with codes are provided too.