SuP: Sub-cloud Driven Point Cloud Registration

SuP: Sub-cloud Driven Point Cloud Registration

Sheldon Fung¹, Wei Pan², Ling Cao², Fei Hou³,⁴, Ling Chen⁵, Shasha Mao⁶, Hongdong Li⁷, Xuequan Lu¹⁎

¹University of Western Australia · ²OPT Machine Vision · ³Institute of Software, CAS · ⁴University of Chinese Academy of Sciences · ⁵University of Technology Sydney · ⁶Xidian University · ⁷Australian National University

⁎ Corresponding author.

📄 Paper: coming soon  |  📦 Code: this repository  |  🗂️ Data (C3DM / C3DLM): Google Drive

---

Introduction

Existing learning-based point-cloud-registration methods handle high-overlap pairs well but struggle when the overlap is low, because geometric or semantic similarities in the non-overlapping regions inevitably produce ambiguous matches.

SuP reformulates low-overlap registration as a high-overlap sub-cloud anchor pair mining problem. The core component is the Dual-phase Sub-cloud Anchor Mining (DSAM) module:

1. Subdivide the source and target point clouds into multiple sub-clouds.
2. Phase 1 — OPS (Overlap-guided Prior-weighting Scheme): leverages feature salience to cheaply pre-score candidate sub-cloud anchor pairs.
3. Phase 2 — MPN (Multi-scale Post-weighting Network): refines the ranking by exploiting neighborhood feature consensus at multiple scales.
4. Merge-to-match: the top-ranked anchor pairs are merged to produce final dense correspondences, from which the transformation is recovered via either LGR (RANSAC-free) or RANSAC.

DSAM is supervised end-to-end by an alignment-aware weighting loss (AWL) that uses on-the-fly anchor-pair alignment errors as the ranking target.

Main results

Registration Recall (%) on Color3DMatch (C3DM) and Color3DLoMatch (C3DLM):

Method	Estimator	C3DM RR ↑	C3DLM RR ↑
CoFiNet	RANSAC-50k	89.3	67.5
GeoTransformer	RANSAC-50k	92.0	75.0
PEAL	RANSAC-50k	94.6	81.7
ColorPCR	RANSAC-50k	96.7	88.9
SuP (ours)	RANSAC-50k	98.1	90.4
CoFiNet	LGR	87.6	64.8
GeoTransformer	LGR	91.5	74.0
PEAL	LGR	94.3	81.2
ColorPCR	LGR	96.5	88.3
SuP (ours)	LGR	97.8	90.2

SuP is the new state of the art on both benchmarks under both estimators; the LGR (RANSAC-free) numbers also match or exceed prior RANSAC-50k baselines.

Installation

SuP is tested with PyTorch ≥ 1.13 + CUDA ≥ 11.7 and builds two custom C++/CUDA extensions for point-cloud subsampling and radius neighbor search.

Ubuntu / Linux

git clone https://github.com/SheldonFung98/SuP.git
cd SuP

# (optional) create a fresh environment
conda create -n sup python=3.10 -y
conda activate sup

# install PyTorch matching your CUDA toolkit (1.13+ recommended)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# project dependencies
pip install -r requirements.txt

# build the C++/CUDA extensions
python setup.py build_ext --inplace

The Docker workflow (./config.sh + ./post_config.sh) used during development is also supported on Linux.

Windows

The repo now compiles cleanly under MSVC (Visual Studio 2019 / 2022). Prerequisites:

• Visual Studio Build Tools 2019 or 2022 with the "Desktop development with C++" workload.
• CUDA Toolkit matching your PyTorch build (e.g. CUDA 11.8 → PyTorch 2.0+cu118).
• Python 3.10 (Anaconda recommended).

From a "x64 Native Tools Command Prompt for VS 2022" or a PowerShell with vcvarsall.bat sourced:

git clone https://github.com/SheldonFung98/SuP.git
cd SuP

conda create -n sup python=3.10 -y
conda activate sup
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# build C++/CUDA extensions (uses MSVC + nvcc; setup.py auto-selects /O2 /std:c++17 /EHsc on Windows)
python setup.py build_ext --inplace

If you see LNK2019 unresolved external symbol from torch_python.lib, make sure the PyTorch wheel matches your Python version and that cl.exe is on PATH (i.e. you launched from the VS Native Tools shell).

Pre-trained weights

Coming soon — will be released here as a GitHub Release.

Color3DMatch / Color3DLoMatch

Data layout

Download the dataset here and arrange it as:

dataset/
├── data/
│   ├── train/7-scenes-chess/cloud_bin_0.npy
│   │   └── ...
│   └── test/7-scenes-redkitchen/cloud_bin_0.npy
│       └── ...

Training

The training entry point lives in experiments/SOAR:

# single GPU
CUDA_VISIBLE_DEVICES=0 python trainval.py

Testing

# 3DMatch
CUDA_VISIBLE_DEVICES=0 ./eval.sh <epoch> <benchmark>

<epoch> is the checkpoint epoch id; <benchmark> is one of 3DMatch, 3DLoMatch, first, second, third, forth. The latter four correspond to overlap bins [0.1–0.15, 0.15–0.2, 0.2–0.25, 0.25–0.3]; the first two cover >0.3 (C3DM) and 0.1–0.3 (C3DLM) respectively.

To evaluate a released checkpoint directly:

CUDA_VISIBLE_DEVICES=0 python test.py --snapshot=../../weights/ckpts.pth.tar --benchmark=3DMatch
CUDA_VISIBLE_DEVICES=0 python eval.py --benchmark=3DMatch --method=lgr

Multi-GPU training

CUDA_VISIBLE_DEVICES=<GPUS> python -m torch.distributed.launch \
    --nproc_per_node=<N_GPU> --master_port=<PORT> trainval.py

# example: 2 GPUs
CUDA_VISIBLE_DEVICES=0,5 python -m torch.distributed.launch \
    --nproc_per_node=2 --master_port=29501 trainval.py

Citation

If you find SuP useful, please cite:

@inproceedings{fung2026sup,
  title     = {SuP: Sub-cloud Driven Point Cloud Registration},
  author    = {Fung, Sheldon and Pan, Wei and Cao, Ling and Hou, Fei and
               Chen, Ling and Mao, Shasha and Li, Hongdong and Lu, Xuequan},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  note      = {Highlight}
}

Acknowledgements

SuP builds on the excellent work of:

• PREDATOR
• CoFiNet
• GeoTransformer
• ColorPCR