Skip to content

Latest commit

 

History

History
22 lines (13 loc) · 1.03 KB

c3dpo.md

File metadata and controls

22 lines (13 loc) · 1.03 KB
Raw

C3DPO: Canonical 3D Pose Networks for Non-Rigid Structure From Motion

December 2019

tl;dr: Infer 3D pose for non-rigid objects by introducing DL to non-rigid structure-from-motion (NR-SFM).

Overall impression

C3DPO transforms closed-formed matrix decomposition problem into a DL-based parameter estimation problem. This method is faster and also can embody prior info that is not apparent in the linear model.

A challenge in NR-SFM is the ambiguity of internal object deformation (or pose in this paper, non-rigid motion) and viewpoint changes (rigid motion). C3DPO introduces a canonicalization network to encourage the consistent decomposition.

Key ideas

  • The main takeaway from this work: Work as many constraints as possible into loss. Use any mathematical cycle-consistency to constrain learning.
  • Use deep learning to supplement maths, not to replace math.

Technical details

  • Summary of technical details

Notes

  • Questions and notes on how to improve/revise the current work