MultiChange3D Dataset

Overview

3D change detection (3DCD) is essential for monitoring infrastructure, environmental dynamics, and natural hazards. Existing algorithms are typically evaluated on single-scene datasets, and their generalization across varied real-world scenes remains largely unexplored due to the absence of a universal benchmark.

MultiChange3D provides co-registered pairs of point clouds with ground-truth geometric change labels across 11 diverse scenes captured by 6 sensor types, enabling standardized evaluation and cross-scene generalization analysis for 3DCD methods.

Representative scenes from the MultiChange3D dataset. The Bike parking construction scene is colorized by LiDAR intensity; the others by RGB values.

Explore the Data

Select a scene to interactively view the co-registered point cloud pair. Both views share the same camera — rotate, pan, and zoom in either panel.

RGB-D Office Open Space (RGB-D)

MLS Open Space (MLS) Underground Car Parking Bike Parking Construction Vineyard

TLS Classroom Meeting Room

UAV Camera Landslide

Airborne City (Camera) City (LiDAR)

Epoch 0

Epoch 1

Dataset

Each scene folder contains co-registered point cloud pairs and ground-truth labels for all available epoch pairs. Point clouds are provided in .ply format.

Sensor	Scene	Approx. Points	Avg. Density (m)	Epochs	Pairs	Extra Features	Condition	Download
RGB-D	Office	2 M	0.002	4	6	RGB	Indoor, cluttered	Link
RGB-D	Open Space (RGB-D)	4 M	0.002	4	6	RGB	Indoor, furniture changes	Link
MLS	Open Space (MLS)	200 k	0.01	2	1	Intensity	Indoor, furniture changes	Link
MLS	Underground Car Parking	24 M	0.01	3	3	Intensity	Indoor, vehicle motion	Link
MLS	Bike Parking Construction	5 M	0.02	4	6	Intensity	Outdoor, construction	Link
MLS	Vineyard*	5 M	0.02	3	3	—	Outdoor, vegetation	Link
TLS	Classroom	40 M	0.005	2	1	Intensity, RGB	Indoor, furniture changes	Link
TLS	Meeting Room	170 M	0.003	2	1	Intensity, RGB	Indoor, small-scale	Link
UAV Camera	Landslide**	20 M	0.04	4	4	RGB	Outdoor, natural terrain	Link
Airborne Camera	City	800 M	0.05	2	1	RGB	Simulated changes, urban	Link
Airborne LiDAR	City	350 M	0.1	2	1	Intensity, RGB	Outdoor, large-scale urban	Link

* No ground-truth change labels. ** Data from Galve et al., 2025.

Benchmark

An initial benchmark of 7 methods from 3 categories, evaluated on three representative scenes: Open Space (RGB-D), Bike Parking Construction, and Landslide.

Euclidean distance-based: C2C, M3C2
3D displacement estimation-based: F2S3, Landslide-3D
Deep learning classification: Siamese KPConv, EF-Siamese KPConv, PGN3DCD

Same-scene evaluation (trained and tested per scene)

Scene	Method	Precision	Recall	F1	OA	mAcc	mIoU
Open Space (RGB-D)	C2C	95.5	62.7	75.4	88.5	80.9	73.2
	M3C2	51.8	55.4	52.2	71.7	67.1	51.1
	F2S3	35.8	99.9	51.7	46.3	61.5	29.5
	Landslide-3D	40.2	99.5	55.9	54.6	67.7	38.0
	SKPConv	87.5	72.5	75.6	90.2	84.0	76.1
	EF-SKPConv	89.0	94.9	91.4	95.8	95.6	89.9
	PGN3DCD	97.8	92.7	95.0	97.4	95.8	93.4
Bike Parking Construction	C2C	51.5	50.1	49.7	78.0	68.2	54.3
	M3C2	26.3	43.6	32.0	61.5	55.3	38.3
	F2S3	29.0	93.3	43.8	49.1	64.8	32.0
	Landslide-3D	28.0	96.7	43.1	46.0	64.0	29.4
	SKPConv	85.8	45.9	59.2	86.3	72.0	63.4
	EF-SKPConv	89.9	60.7	72.1	89.9	79.6	72.5
	PGN3DCD	94.0	65.3	76.8	91.5	82.1	76.2
Landslide	C2C	85.5	91.9	88.4	95.6	94.2	87.2
	M3C2	83.8	84.6	84.2	94.5	90.6	83.1
	F2S3	31.2	98.1	47.3	62.1	76.3	42.7
	Landslide-3D	37.1	95.8	53.4	71.0	80.8	50.9
	SKPConv	97.0	93.7	95.3	98.4	96.5	94.5
	EF-SKPConv	97.8	90.6	94.0	98.0	95.1	93.2
	PGN3DCD	96.1	92.4	94.1	98.0	95.8	93.3

Key Findings

Deep learning methods achieve the best scores when trained and tested on the same scene.
Cross-scene generalization remains a significant challenge for learned approaches (OA drops of 16–65 pp).
Euclidean distance-based methods (C2C, M3C2) show stable performance across diverse scenes.
3D displacement estimation-based methods can detect large changes but struggle with planar structures.

Qualitative Results

Open Space (RGB-D), epoch pair 0-2.

Bike Parking Construction, epoch pair 0-2.

Landslide, epoch pair 0-1.

Citation

Published at ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, proceedings of the ISPRS Congress 2026.

@article{wang2026multichange3d, title = {{MultiChange3D}: A Multi-Scene, Multi-Sensor Dataset for Benchmarking 3D Geometric Change Detection}, author = {Wang, Zhaoyi and Tryba{\l}a, Pawe{\l} and Wieser, Andreas and Remondino, Fabio}, journal = {ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci.}, year = {2026}, volume = {XI-2-2026}, pages = {639--646}, URL = {https://isprs-annals.copernicus.org/articles/XI-2-2026/639/2026/}, DOI = {10.5194/isprs-annals-XI-2-2026-639-2026} }

License

The data is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. The code is licensed under the MIT License.