RIGI: Rectifying Image-to-3D Generation Inconsistency via Uncertainty-Aware Learning.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Image-to-3D generation seeks to create 3D models from single 2D images.
• Conventional methods often fail due to epistemic uncertainty from imperfect multi-view data.
• Deterministic optimization in existing pipelines neglects crucial uncertainty information.

Purpose of the Study:

• To develop an uncertainty-aware optimization framework for more robust and reliable 3D generation.
• To explicitly model and mitigate epistemic uncertainty in the image-to-3D pipeline.
• To enhance the geometric and perceptual plausibility of generated 3D models.

Main Methods:

• Implemented a progressive sampling strategy with varying camera elevations to enhance viewpoint coverage.
• Estimated epistemic uncertainty using discrepancies between two independently optimized Gaussian models.
• Incorporated an uncertainty map into regularization for adaptive loss weighting, suppressing unreliable supervision.

Main Results:

• Significantly reduced artifacts and inconsistencies in generated 3D models.
• Demonstrated higher-quality 3D generation compared to conventional approaches.
• Provided theoretical analysis with a probabilistic upper bound on expected generation error.

Conclusions:

• The proposed uncertainty-aware optimization framework enhances image-to-3D generation robustness.
• Explicitly modeling epistemic uncertainty leads to superior 3D model quality.
• The method offers a promising direction for reliable 3D reconstruction from single images.