Dynamic Non-Rigid Objects Reconstruction with a Single RGB-D Sensor.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Dynamic non-rigid object reconstruction is challenging due to error accumulation and surface tracking failures in sequential methods.
• Existing methods struggle with long sequences and maintaining accuracy for deformable objects.

Purpose of the Study:

• To develop a robust global non-rigid registration framework for accurate 3D reconstruction of dynamic objects.
• To overcome limitations of sequential fusion and surface tracking in 3D reconstruction tasks.
• To achieve high-fidelity geometry and appearance reconstruction for deformable objects.

Main Methods:

• A novel global non-rigid registration framework is proposed, incorporating fusion of partial scans.
• Pairwise non-rigid registration and loop detection are employed to establish correspondences.
• A global registration procedure aligns pieces, followed by a model-update step for refinement.
• Both geometric and appearance constraints are enforced during alignment.

Main Results:

• The framework successfully reconstructs complete and watertight deformable objects.
• Accurate geometry and high-fidelity color maps are achieved for the reconstructed mesh.
• Experiments on synthetic and real datasets validate the approach's effectiveness.

Conclusions:

• The proposed global registration framework effectively addresses challenges in dynamic non-rigid 3D reconstruction.
• The explicit loop closure mechanism mitigates error accumulation and improves robustness.
• This method enables high-quality reconstruction of complex deformable objects from single RGB-D sensor data.