Original title: Stable and Scalable Bundle Adjustment of Holistic 3D Structures
Shaohui Liu, Rémi Pautrat, Daniel Barath, Richard Hartley +2 · September 2026
Computers can build 3D digital scenes from flat photos, but matching extra real-world shapes like parallel lines or flat surfaces usually makes the math unstable and painfully slow. Researchers created a unified system to solve this issue. They organized shapes into two buckets: simple pieces like points or lines that show up directly in images, and structural groups like parallel lines. Next, they treated those structural groups just like extra cameras inside the math. This allowed them to measure all the grouping rules and connections using standard 2D image predictions instead of tricky 3D formulas. Because of this trick, the math keeps the clean, fast pattern of standard point-based tools. Tests on real and synthetic data showed the method matches the speed of classic point-only systems while delivering much richer shapes and better geometric accuracy.
Imagine packing a suitcase where each outfit must match others, but the space is limited. Choosing one item affects which others you can take, and planning combinations takes extra time. In this paper, the challenge is adding complex 3D relations without slowing down modeling, and the solution is to treat these relations as if they were single items that fit automatically, making the whole process efficient.
Engineers building 3D models for autonomous vehicles or virtual reality can now include complex structures like parallelism without slowdowns. This makes the models more accurate and reliable for real-world use.
The abstract does not specify limitations, such as the types of geometric structures that may not work or scalability to extremely large datasets beyond those tested.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Stable and Scalable Bundle Adjustment of Holistic 3D Structures — Shaohui Liu, Rémi Pautrat, Daniel Barath et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.