PARTE

Benchmark data

Datasets

To reduce download size and provide consistent registration inputs, our dataset package contains voxel-downsampled point clouds together with the benchmark pair definitions used in the evaluation.

Download processed datasetsZIP · 1.15 GiB
SHA-256 checksum
e117ddcbcb11c4ca666ba26c3f053d43f9407e88ef02e42c63442f6590480d4d
Dataset layout
processed_datasets/
├── scans/<dataset>/<sequence>/
│   ├── cloud_bin_<index>.ply
│   └── voxel_size
├── benchmarks/<dataset>/<sequence>/
│   └── gt.log
├── indoor.txt
├── outdoor.txt
└── additional.txt

Original data sources

3DMatch and 3DLoMatch
Point clouds and ground-truth evaluation data are available from the 3DMatch project.
ETH
Download the scans from the ETH Laser Registration Dataset and the benchmark pair data from the 3DSmoothNet repository.
KITTI-10m and KITTI-LC
Download the scans from the KITTI Odometry Benchmark. We use the ICP-refined ground truths provided by the LiDAR Registration Benchmark.
RESSO P2F
The original scans are available from the RESSO project page.

PARTEPlane-Assisted Robust Transformation Estimationfor Point Cloud Registration

PaperCodeVideo

A global registration method that treats planar structure as complementary evidence, combining point and plane correspondences in confidence-weighted graph-based outlier rejection.

Abstract

Global point-cloud registration remains challenging when limited overlap, repetitive geometry, and sensor noise produce correspondence sets dominated by outliers. Planar regions are particularly difficult for conventional point descriptors and are therefore often suppressed or discarded before matching. PARTE instead represents planar structure as complementary registration evidence using Plane Context Histograms, and combines candidate point and plane correspondences in a confidence-weighted compatibility graph for joint outlier rejection and rigid transformation estimation. Evaluation covers 8,097 pairs from six indoor and outdoor benchmarks spanning dense RGB-D and sparse LiDAR measurements.

Pipeline

Explore the complete PARTE registration pipeline.

Select any dataset, sequence, and pair from the complete evaluation set. Use the interactive visualizer below to examine each stage of registration.

Reported Benchmark Results

Loading example…

Loading the voxel-downsampled benchmark pair…

Stage 1 of 6InputInput point-cloud pair.
Loading interactive 3D visualizationThree.js is loaded only when this section enters the viewport.
Target scanSource scan
Planar patches
Target non-planar pointsSource non-planar pointsAll point correspondences
All PCH plane matches
Selected plane matchesSelected point matches
Selected plane matchesSelected point matches
All point correspondencesSelected point matches
Selected point matches
Selected point matches
Scans

Evaluation

Evaluation across dense RGB-D and sparse LiDAR benchmarks.

The evaluation contains 8,097 registration pairs from six indoor and outdoor datasets. The comparison below uses the same evaluation criteria and end-to-end timing policy as the paper.

Indoor RGB-D89.6 / 55.8%3DMatch / 3DLoMatch

Second-highest success rate on both benchmarks.

Outdoor odometry99.8%KITTI-10m

Tied for the highest overall success rate.

Partial-to-full59.6%RESSO P2F

Highest overall success rate.

Additional LiDAR99.6%ETH

Highest overall success rate.

Registration comparisonSuccess rate and end-to-end runtimeEach cell reports success rate (%) above mean runtime (ms).
Method3DMatch3DLoMatchKITTI-10mKITTI-LCRESSO P2FETH
RANSAC-10⁵Baseline85.670.8 ms45.870.3 ms91.5416.7 ms42.6376.9 ms34.3358.1 ms63.3892.9 ms
KISS-MatcherNative pipeline75.5110.8 ms26.826.4 ms97.198.0 ms52.1101.1 ms42.4213.5 ms89.8226.7 ms
CLIPPER+Graph-based85.9117.5 ms44.7130.1 ms98.7974.1 ms65.42574.9 ms58.6707.9 ms94.54266.4 ms
TEASER++Graph-based86.677.7 ms48.064.5 ms98.9519.8 ms70.4494.7 ms56.6385.2 ms94.01078.3 ms
MACGraph-based88.0112.4 ms52.485.3 ms98.41908.7 ms66.21721.4 ms57.6549.9 ms87.27521.2 ms
VBRegLearning-based89.4102.8 ms53.594.0 ms97.5496.4 ms62.9465.1 ms57.6443.7 ms90.51015.4 ms
PREDATORLearning-based92.195.7 ms61.082.0 ms99.8553.3 ms28.8366.3 ms10.1392.3 ms64.01657.8 ms
G3RegPlane-based4.835.1 ms0.042.9 ms99.188.0 ms86.296.5 ms——32.7161.9 ms
PARTEOurs89.659.4 ms55.856.8 ms99.8178.0 ms85.8127.1 ms59.6315.8 ms99.6437.6 ms

Bold and underlined values indicate the best and second-best success rates in the complete paper tables.

Evaluation protocolBenchmark datasets and success criteriaPoint counts are measured after voxel downsampling.
DatasetSensorScenesPairsvPts / scanPlanes / scanSuccess criterion
3DMatchRGB-D81,6235 cm4,8384.6≤ 15°, ≤ 30 cm
3DLoMatchRGB-D81,7815 cm4,8214.6≤ 15°, ≤ 30 cm
KITTI-10mLiDAR355630 cm18,2385.1≤ 5°, ≤ 0.6 m
KITTI-LCLiDAR53,32530 cm18,3547.9≤ 5°, ≤ 2 m
RESSO P2FTLS + RGB-D7995 cm10,1437.4≤ 15°, ≤ 30 cm
ETHLiDAR471310 cm30,2342.0≤ 15°, ≤ 30 cm