NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
CLONE: A 3DGS-Based Closed-Loop Differentiable Optimization Framework for Single-Image Normal Estimation
TL;DR
Single-image normal estimation, which determines the surface orientation from a single image, often struggles with limited supervision and geometric constraints. CLONE is a 3DGS-based Closed-Loop differentiable Optimization framework that integrates image geometry into the estimation process.
✦ Why It Matters
Engineers can leverage CLONE to improve normal estimation in applications like 3D modeling and computer vision.
Key Takeaways
- Utilize 3D Gaussian Splatting for scene parameterization.
- Implement a differentiable illumination model for better image-to-geometry mapping.
- Incorporate a refinement network to enhance local geometric details.
- Optimize under a unified reprojection objective for improved accuracy.
How It Works
CLONE operates by creating an 'image-geometry-image' consistency loop. It first uses 3D Gaussian Splatting to derive continuous surface normals through covariance eigen-decomposition, which provides a clear gradient pathway for optimization.
The differentiable illumination model links surface normals to image radiance, allowing reprojection errors to guide the optimization. A refinement network enhances local details while maintaining differentiability, and a gating mechanism balances global consistency with local detail.