Robust Isometric Non-Rigid Structure-From-Motion.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Modeling

Background:

• Non-Rigid Structure-from-Motion (NRSfM) reconstructs 3D objects from 2D images.
• Current NRSfM methods struggle with correspondence errors, limiting their practical application.
• Automatic correspondence establishment is prone to errors, hindering NRSfM scope.

Purpose of the Study:

• To develop a statistically robust NRSfM pipeline.
• To overcome limitations posed by erroneous keypoint correspondences.
• To enable reliable 3D reconstruction from automatically established correspondences.

Main Methods:

• A three-step automatic pipeline exploiting isometry for robust NRSfM.
• Step (i): Optical flow computation with robustification via warp estimation.
• Step (ii): Local normal reconstruction and surface integration.
• Step (iii): Rejection of points violating local isometry using a novel scale-independent measure.

Main Results:

• The proposed method effectively handles and discards erroneous correspondences.
• Demonstrated robust optical flow estimation and local normal reconstruction.
• Achieved superior performance compared to existing NRSfM methods on synthetic and real datasets.

Conclusions:

• The developed NRSfM pipeline significantly improves robustness against correspondence errors.
• Exploiting isometry provides a powerful mechanism for error detection and rejection.
• The method broadens the applicability of NRSfM in computer vision and robotics.