IncreLM: Incremental 3D Line Mapping


Xulong Bai (Institute of automation, Chinese academy of science, Chinese Academy of Sciences), Hainan Cui (Chinese Academy of Sciences), Shuhan Shen (Institute of automation, Chinese academy of science)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Given posed images and Structure-from-Motion (SfM) points, we aim to produce 3D line segments (3D LSs) and line tracks. Traditional methods typically reconstruct a single 3D LS corresponding to each 2D line segment (2D LS) independently and in parallel, later merging all single 3D LSs to produce final 3D LSs and line tracks. However, this independence may lead to inconsistencies in outlier line matches identified across different single 3D LS reconstructions. To enhance the robustness of outliers, we propose an incremental 3D line mapping method that sequentially reconstructs each final 3D LS, using outlier line matches identified from earlier reconstructions to guide later ones. In our approach, 3D LS hypotheses are generated through two-view line triangulation, utilizing 3D points and vanishing points within a hybrid RANSAC framework. A graph is then created, with nodes representing the hypotheses and edges linking nodes that share identical source 2D LSs. Initially, the best 3D LS hypothesis is found based on neighborhood supports and added to an empty 3D line map. We then extend the line track of the new 3D LS and filter out outlier matches through reprojection. Next, we filter out outlier nodes generated by outlier matches, locate the next-best 3D LS hypothesis, and integrate it into the existing map. This iterative process continues until no best 3D LS hypotheses can be identified from the graph. Finally, the line tracks are merged and a joint optimization is performed to improve the map quality. Experiments show that our system exceeds current state-of-the-art methods in completeness and accuracy and produces longer line tracks. Code is available at https://github.com/3dv-casia/IncreLM.

Citation

@inproceedings{Bai_2024_BMVC,
author    = {Xulong Bai and Hainan Cui and Shuhan Shen},
title     = {IncreLM: Incremental 3D Line Mapping},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0256.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection