Multi-view Homographies from Dense Matches
(2025) In Master’s Theses in Mathematical Sciences FMAM05 20251Mathematics (Faculty of Engineering)
- Abstract
- We propose a SLAM system that reconstructs 3D scenes by leveraging dense feature correspondences and planar structures, identified through multi-view homographies. In contrast to traditional SLAM algorithms, our method emphasizes planes as the primary geometric primitive.
While classical keypoint detectors such as SIFT typically yield a few hundred matches per image pair, modern dense matchers can produce on the order of 10,000 matches. The abundance of correspondences offers potential improvements in performance for many computer vision tasks, but also calls for new algorithms to more efficiently utilize them. A recent paper from our research group introduced a method for clustering and summarizing dense matches for essential matrix... (More) - We propose a SLAM system that reconstructs 3D scenes by leveraging dense feature correspondences and planar structures, identified through multi-view homographies. In contrast to traditional SLAM algorithms, our method emphasizes planes as the primary geometric primitive.
While classical keypoint detectors such as SIFT typically yield a few hundred matches per image pair, modern dense matchers can produce on the order of 10,000 matches. The abundance of correspondences offers potential improvements in performance for many computer vision tasks, but also calls for new algorithms to more efficiently utilize them. A recent paper from our research group introduced a method for clustering and summarizing dense matches for essential matrix estimation. This thesis adapts that technique to the estimation of plane normals, including local optimization and bundle adjustment. Our formulation reduces computational cost without sacrificing the quality of the reconstruction.
Although our system processes dense matches efficiently, it does not yet match the accuracy of state-of-the-art SLAM pipelines. Still, it demonstrates the potential of plane-based representations for robust and efficient scene reconstruction. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9205329
- author
- Nilsson Gisleskog, Gustav LU
- supervisor
- organization
- course
- FMAM05 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- publication/series
- Master’s Theses in Mathematical Sciences
- report number
- LUTFMA-3589-2025
- ISSN
- 1404-6342
- other publication id
- 2025:E41
- language
- English
- id
- 9205329
- date added to LUP
- 2025-09-15 11:10:58
- date last changed
- 2025-09-15 11:10:58
@misc{9205329, abstract = {{We propose a SLAM system that reconstructs 3D scenes by leveraging dense feature correspondences and planar structures, identified through multi-view homographies. In contrast to traditional SLAM algorithms, our method emphasizes planes as the primary geometric primitive. While classical keypoint detectors such as SIFT typically yield a few hundred matches per image pair, modern dense matchers can produce on the order of 10,000 matches. The abundance of correspondences offers potential improvements in performance for many computer vision tasks, but also calls for new algorithms to more efficiently utilize them. A recent paper from our research group introduced a method for clustering and summarizing dense matches for essential matrix estimation. This thesis adapts that technique to the estimation of plane normals, including local optimization and bundle adjustment. Our formulation reduces computational cost without sacrificing the quality of the reconstruction. Although our system processes dense matches efficiently, it does not yet match the accuracy of state-of-the-art SLAM pipelines. Still, it demonstrates the potential of plane-based representations for robust and efficient scene reconstruction.}}, author = {{Nilsson Gisleskog, Gustav}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master’s Theses in Mathematical Sciences}}, title = {{Multi-view Homographies from Dense Matches}}, year = {{2025}}, }