💡 The idea of Part-Hierarchy Learning

Part-Hierarchy Learning aims to understand objects/ concepts by parsing them into parts hierarchically like parse tree models, and my technical route is: (i) finding part representation and transferring them into the representation space of object whole; (ii) aligning part poses for aggregation to present whole; (iii) merging parts to represent whole. The core question is how to organize part representation for better depicting whole and how to implant a/ several parse tree(s) into one neural network.