Intermediate
(1) 2D image -> Intermediate
(Real image as input)
(2) Intermediate -> Specific kind of 3D shape
(Synthetic data as input)
(3) Reprojection: Estimated 3D shape -> Intermediate
(Synthetic data, with ground truth 3D shape data)
Intermediate: Use Intrinsic object properties
相較於傳統重建上,這裡多了一層 Intermediate 來聯繫 Input dataset & predicted output,而由於作為 Intermediate 的 Intrinsic properties 為無關環境因素,純粹與物體本身有關的內部性質,如深度、法向量、骨架等,因此以此作為 input 訓練的結果,便不具有 Domain issue 的問題。