Reproduce “Fine Dense Alignment of Image Bursts through Camera Pose and Depth Estimation” on OpenResearch

About this paper

Bruno Lecouat, Yann Dubois de Mont-Marin, Théo Bodrito, Julien Mairal, Jean Ponce

Researchers developed a non-learning, optimization-based framework for fine dense alignment of image bursts from handheld cameras, jointly estimating camera poses and scene depth. The method achieved lower optical flow error and higher accuracy for pose and depth estimation on synthetic datasets, and enabled superior noise reduction in real-world night photography compared to existing methods.

Key insights

• Accurately leveraging the 3D scene structure and camera motion is essential for fine alignment in burst photography, particularly where traditional 2D methods struggle with parallax.
• A training-free, global optimization framework, when geometrically principled, can surpass data-driven learning methods in dense optical flow, camera pose, and depth estimation for small-motion bursts.
• Parameterizing dense optical flow via a shared 3D scene model (depth and normals) and explicit camera poses ensures multi-view consistency and superior accuracy compared to pairwise 2D flow estimations.

Summary by alphaXiv.