gradient-spaces/Rectified-Point-Flow

[NeurIPS'25] Rectified Point Flow: Generic Point Cloud Pose Estimation

Rectified Point Flow (RPF) is a unified model that formulates pairwise point cloud registration and multi-part shape assembly as a single conditional generative problem. Given unposed point clouds, the method learns a continuous point-wise velocity field that transports noisy points toward their target positions, from which part poses are recovered.

Installation

git clone https://github.com/GradientSpaces/Rectified-Point-Flow.git
cd Rectified-Point-Flow
conda create -n py310-rpf python=3.10 -y
conda activate py310-rpf
poetry install  # or `uv sync`, `bash install.sh`

Quick Start

# Assembly Generation:
python sample.py data_root=./demo/data

# Overlap Prediction:
python predict_overlap.py data_root=./demo/data

More details can be found in our GitHub Repo.

Checkpoints

• RPF_base_full_*.ckpt: Complete model checkpoint for assembly generation
• RPF_base_pretrain_*.ckpt: Encoder-only checkpoint for overlap prediction

Training Data

Dataset	Task	Part segmentation source	Parts per sample
IKEA-Manual	Shape Assembly	Defined by IKEA manuals	[2, 19]
PartNet	Shape Assembly	Human-annotated parts	[2, 64]
BreakingBad-Everyday	Shape Assembly	Simulated fractures via fracture-modes	[2, 49]
Two-by-Two	Shape Assembly	Annotated by human	2
ModelNet-40	Pairwise Registration	Following Predator split	2
TUD-L	Pairwise Registration	Real scans with partial observations	2
Objverse	Overlap Prediction	Segmented by SAMPart3D	[3, 12]

Citation

@inproceedings{sun2025_rpf,
    author = {Sun, Tao and Zhu, Liyuan and Huang, Shengyu and Song, Shuran and Armeni, Iro},
    title = {Rectified Point Flow: Generic Point Cloud Pose Estimation},
    booktitle = {NeurIPS},
    year = {2025},
}