Reproduce “SuperPrimitive: Scene Reconstruction at a Primitive Level” on OpenResearch

About this paper

Kirill Mazur, Gwangbin Bae, Andrew J. Davison

This work from the Dyson Robotics Lab at Imperial College London introduces SuperPrimitive, a novel intermediate representation for 3D scene reconstruction that effectively combines learned local geometric priors with multi-view optimization. The approach improves robustness and efficiency in tasks such as sparse depth completion, few-view Structure-from-Motion, and monocular visual odometry, achieving strong performance across various benchmarks.

Key insights

• Leveraging neural network predictions for segmentation and surface normals allows robust estimation of unscaled local geometry within image regions, forming SuperPrimitives.
• By optimizing a single scale factor per SuperPrimitive instead of per-pixel depths, the dimensionality of the multi-view optimization problem is significantly reduced, improving efficiency and robustness.
• Combining learned local geometric priors (from neural networks) with traditional multi-view photometric optimization effectively addresses ambiguities and ensures global consistency.

Summary by alphaXiv.