SCORE: Saturated Consensus Relocalization in Semantic Line Maps

Our paper has been accepted by IROS 2025, and a longer version can be found on arxiv: https://arxiv.org/pdf/2503.03254 2 hour video in Chinese: https://www.bilibili.com/video/BV1GPnhzCE5N/?spm_id_from=333.337.search-card.all.click

1. Test our matlab and CPP implementation

• The test codes use data under csv_dataset/test_data, available after git clone.
• matlab/test.m (check first your platform can use our compiled mex functions).
• CPP/test.cpp (compile first, require Eigen3 and OpenMP).

cd CPP && mkdir build && cd build && cmake .. && make

2. Download our datasets

including

• preprocessed csv datasets (CSV_dataset.tar.gz)
• semantic segmentation masks (*_segmentation_mask.tar.gz).

Google drive link: https://drive.google.com/drive/folders/141lQdHufOMp3ovRSQsV_06HDgxwhTKuF?usp=sharing

3. Replicate experiment results presented in our paper:

• Extract CSV_dataset.tar.gz downloaded from Google drive and move files (S1 S1_pred ...) into folder csv_dataset.
• Choose any query image and relocalize it with CPP/reloc_one.cpp.
• Run matlab/Experiments/rotation.m, pipeline.m, and sensitivity_plot.m in sequence.
• The matlab codes take time, we recommend a fast trial in scene S2 office and use gt semantic labels.
• Refer to README files under corresponding folders for more details.

Disclaimer

Since the publication of the paper, we have been continuously improving the codebase. As a consequence, the results might slightly deviate from (usually improve over) the original numbers found in our IROS2025 paper.

4. Go through line map construction pipeline on ScanNet++ Dataset

Download ScanNet++ Dataset

We revised the pre-processing code for ScanNet++. Follow README.md under folder scannetpp.

Semantic Segmentation Pipeline combing RAM++ and Grounded-SAM

Follow README.md under folder semantic_pipeline.

Extract 3D Semantic Line Maps

Follow README.md under folder line_map_extractor.

Citation

If you find our work helpful, please cite:

@article{jiang2025score,
  title={SCORE: Saturated Consensus Relocalization in Semantic Line Maps},
  author={Jiang, Haodong and Zheng, Xiang and Zhang, Yanglin and Zeng, Qingcheng and Li, Yiqian and Hong, Ziyang and Wu, Junfeng},
  journal={arXiv preprint arXiv:2503.03254},
  year={2025}
}