Procrustean Normal Distribution for Non-Rigid Structure from Motion.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Non-rigid structure from motion (NRSfM) is crucial for analyzing deformable objects.
• Existing NRSfM methods often impose restrictive assumptions like fixed-rank or smooth deformations, which are unrealistic.
• Rotation estimation errors can erroneously be interpreted as deformations, impacting accuracy.

Purpose of the Study:

• To develop a novel deformation model for NRSfM that overcomes limitations of existing approaches.
• To introduce a method that strictly separates rigid and non-rigid motion components.
• To eliminate the need for user-defined parameters in NRSfM.

Main Methods:

• Proposed a new prior distribution, the Procrustean normal distribution (PND), specifically for non-rigid deformations.
• Leveraged the PND's ability to represent non-rigid deformations in a linear subspace without strong constraints.
• Developed the EM-PND algorithm to fit the PND to 2D observations for solving NRSfM.

Main Results:

• The PND effectively separates rigid and non-rigid motion, leading to improved performance.
• The EM-PND algorithm achieved state-of-the-art results on benchmark datasets.
• The proposed method requires no user-determined parameters, simplifying the NRSfM process.

Conclusions:

• The Procrustean normal distribution provides a more accurate and flexible deformation model for NRSfM.
• The EM-PND algorithm demonstrates superior performance and robustness in non-rigid structure from motion tasks.
• This work advances NRSfM by offering a parameter-free approach with enhanced deformation modeling capabilities.