A Multi-Scene, Multi-Sensor Dataset for Benchmarking 3D Geometric Change Detection

Zhaoyi Wang¹  |  Paweł Trybała²  |  Andreas Wieser¹  |  Fabio Remondino²

¹ Chair of Geosensors and Engineering Geodesy (GSEG), ETH Zurich, Switzerland
² 3D Optical Metrology (3DOM) Unit, Bruno Kessler Foundation (FBK), Trento, Italy

3D change detection (3DCD) is essential for monitoring infrastructure, environmental dynamics, and natural hazards. However, existing algorithms are often evaluated on single-scene datasets, and their generalization across varied real-world scenes remains largely unexplored.

MultiChange3D is a multi-scene, multi-sensor 3D change detection dataset for identifying geometric changes in 3D space. The dataset provides co-registered pairs of point clouds with ground-truth geometric change labels, enabling standardized evaluation across different methods.

The dataset covers:

• 5 sensor types: RGB-D, MLS, TLS, UAV Camera, Airborne LiDAR, and Airborne Camera
• 10 scenes spanning indoor, outdoor, and city-scale environments
• 2 to 4 epochs per scene, with both natural and induced changes
• Ground-truth annotations per point cloud pair: unchanged, removed, added

MultiChange3D is joint work between ETH Zurich (GSEG) and Bruno Kessler Foundation (3DOM-FBK).

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.

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

The evaluation code, including metric computation, is available in evaluation/. Additional helper scripts are provided in scripts/.

An initial benchmark is provided for 7 methods from 3 categories, evaluated on three representative scenes: Open space (RGB-D), Bike parking construction, and Landslide.

Methods evaluated:

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

Qualitative results on the Open space (RGB-D) scene (epoch pair 0-2).

Qualitative results on the Bike parking construction scene (epoch pair 0-2).

Qualitative results on the Landslide scene (epoch pair 0-1).

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 learning-based approaches (OA drops of 16--65 pp).
• Euclidean distance-based methods (C2C, M3C2) show stable performance across scenes, but are weak in detecting overlapping changes and lack context understanding.

We sincerely acknowledge the following open-source projects for supporting our evaluation:

• F2S3: A geometry-based method for 3D displacement estimation.
• Landslide-3D: A multi-modal method combining 3D geometry and RGB information; only the geometry-based component is used in our evaluation.
• Siamese KPConv: A deep learning-based method for 3DCD.
• EF-Siamese KPConv: An enhanced variant of Siamese KPConv.
• PGN3DCD: A deep learning-based method for 3DCD.

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

If you use this dataset in your research, please cite:

@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}
}

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

The code in this repository is licensed under the MIT License.