Rough or Noisy? Metrics for Noise Estimation in SfM Reconstructions.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Structure from Motion (SfM) generates detailed 3D models but struggles to distinguish surface roughness from noise.
• Current SfM noise reduction methods are suboptimal for precise surface inspection.
• Existing noise algorithms do not leverage SfM's inherent data.

Purpose of the Study:

• To develop novel geometrical and statistical metrics for assessing noise in SfM reconstructions.
• To evaluate the effectiveness of supervised learning models trained with these metrics for noise-roughness discrimination.
• To improve the accuracy of 3D surface inspection by separating true surface features from reconstruction artifacts.

Main Methods:

• Proposed geometrical and statistical metrics for noise assessment.
• Utilized supervised learning models trained on these metrics.
• Incorporated data from both the reconstructed object and the camera setup.
• Created and released a new image dataset for SfM with ground truth annotations.

Main Results:

• Developed metrics that correlate with noise presence on reconstructed surfaces.
• Achieved over 85% accuracy in distinguishing noise from roughness using supervised learning.
• Demonstrated an additional 5-6% performance improvement by including camera setup metrics.
• The proposed solution integrates seamlessly into existing SfM workflows.

Conclusions:

• Novel metrics and supervised learning effectively separate noise from surface roughness in SfM.
• The method enhances the reliability of 3D reconstructions for surface inspection.
• The approach is practical, requiring no additional data or sensors and is dataset-available.