Source-linked AI summary
Paris-Lille-3D: a large and high-quality ground truth urban point cloud dataset for automatic segmentation and classification
Xavier Roynard, Jean-Emmanuel Deschaud, François Goulette
TL;DR
Urban 3D point-cloud learning needs datasets with sufficient quantity, class coverage, and segmentation quality, yet comparable urban datasets do not combine these advantages. The paper introduces Paris-Lille-3D, acquired by mobile laser scanning and hand-segmented and classified, and reports initial automatic segmentation and classification results. The dataset contains 140 million points over 2 km in two cities and assigns objects to 50 classes.
Problem
Urban 3D point-cloud datasets suitable for machine-learning segmentation and classification remain difficult to find, despite increasing data needs.
Method
The paper constructs a Mobile Laser Scanning dataset through trajectory processing, point-cloud generation, hand segmentation, and class labeling.
Results
The dataset contains 140 million points over 2 km in two cities, with hand-segmented objects assigned among 50 classes, and the paper tests an automatic segmentation and classification method.
Takeaways & Limitations
Paris-Lille-3D is intended for classification learning and, because objects are sufficiently separated, for precise segmentation learning.
Takeaways & Limitations
The segmentation method can merge nearby cars or buildings connected by a cable because it defines objects through connected components after ground removal.
Abstract
from arXiv · showhide
This paper introduces a new Urban Point Cloud Dataset for Automatic Segmentation and Classification acquired by Mobile Laser Scanning (MLS). We describe how the dataset is obtained from acquisition to post-processing and labeling. This dataset can be used to learn classification algorithm, however, given that a great attention has been paid to the split between the different objects, this dataset can also be used to learn the segmentation. The dataset consists of around 2km of MLS point cloud acquired in two cities. The number of points and range of classes make us consider that it can be used to train Deep-Learning methods. Besides we show some results of automatic segmentation and classification. The dataset is available at: http://caor-mines-paristech.fr/fr/paris-lille-3d-dataset/
1 Introduction
Paris-Lille-3D addresses the shortage of segmented and classified urban 3D point-cloud datasets by introducing a dataset designed for both precise segmentation learning and classification. The paper also reports initial automatic segmentation and classification results.
- Urban 3D point-cloud datasets remain difficult to find despite the growing need for data supporting machine-learning segmentation and classification.
- The dataset is presented as combining advantages that the authors believe existing comparable datasets do not combine.
- Paris-Lille-3D provides objects segmented sufficiently precisely for segmentation to be learned.
- The paper reports results from automatic segmentation and classification performed on the dataset.
2 Available Datasets
Existing urban 3D point-cloud datasets differ in acquisition, density, occlusion, annotation quality, and class coverage. The paper compares these characteristics across prior datasets and motivates the need for a more suitable benchmark.
- Urban 3D point-cloud datasets support learned classification and segmentation, but vary substantially in data type, quantity, quality, class count, and task coverage.
- Mobile Laser Scanning requires an accurately known 6D vehicle pose for cloud registration, whereas Terrestrial Laser Scanning requires repeated static acquisitions and registration.
- Airborne Laser Scanning does not provide sufficient point density because of acquisition distance and angle.
- Existing urban datasets have heterogeneous strengths and weaknesses that the paper compares quantitatively with its own dataset.
- Oakland 3-D Point Cloud Dataset: The Oakland dataset has low density, few classes, and large shadows behind trees because it uses a side-looking monofiber LiDAR.
- Semantic3D: Semantic3D contains registered clouds with occlusions and point density that depends on distance to the LiDAR.
- Paris-rue-Madame Database: Rue-Madame retains noisy ground-truth annotations, including parts of cars labeled as road.
- IQmulus & TerraMobilita Contest: iQmulus/TerraMobilita uses a monofiber LiDAR that creates shadows behind objects, while wall points behind cars can be classified as car.
3 Our Dataset: Paris-Lille-3D
Paris-Lille-3D is acquired with the L3D2 mobile laser-scanning system across three trajectories in Lille and Paris, then processed into dense, labeled point-cloud files. Its annotations reuse and extend an existing class tree while preserving object-level labels and point attributes.
- 3.1 Acquisition: L3D2 combines GPS, IMU, and a rear-mounted Velodyne HDL-32E LiDAR angled 30 degrees from horizontal.The prototype is installed on a Citroën Jumper.
- 3.1 Acquisition: Trajectory processing uses tightly coupled GPS-RTK/INS Kalman smoothing, coordinate conversion, timestamp interpolation, and a 20 m range filter.Each retained point is represented by position, sensor origin, timestamp, and LiDAR-return intensity.
- 3.2 Description of point clouds: The dataset contains three parts: two trajectories in the Lille agglomeration and one trajectory in Paris.The Lille trajectories are shown separately from the Paris trajectory in Figure 7.
- 3.2 Description of point clouds: The clouds contain 1000–2000 points per square meter on the ground, with anisotropic patterns caused by the multi-beam LiDAR sensor.The density and acquisition pattern are illustrated in Figure 8.
- 3.3 Labeling: CloudCompare3 was used for manual segmentation and classification, with labels based on a reused and extended iQmulus/Terramobilita class tree.Added classes include bicycle racks, statues, distribution boxes, lighting consoles, and windmills; vehicle subclasses distinguish parked, stopped, and moving vehicles.
- 3.3 Labeling: The dataset is distributed in separate PLY files, whose points include coordinates, LiDAR origin, GPS time, reflectance, object labels, and object classes.Separate annotation text files record sample, class, and additional annotation fields.
4 Results of automatic segmentation and classification
The evaluation applies a connectivity-based segmentation and Random-Forest classification pipeline, with improvements to ground extraction and contextual object descriptors. Results expose segmentation errors while indicating that simpler descriptor combinations can provide effective, faster classification.
- Method: The evaluated pipeline extracts ground by region growing, segments remaining points by connectivity, computes object descriptors, and classifies objects with a Random Forest.Descriptors include geometric features, CVFH, GRSD, and ESF.
- Improvements: Two improvements target robustness: better ground-region-growing seeds for segmentation and contextual descriptors for classification.The contextual descriptor measures an object's minimum elevation relative to the detected ground beneath it.
- Segmentation: The segmentation evaluation uses object detection precision and recall under the iQmulus/TerraMobilita matching metric.The supplied passages identify the metric and evaluation setup but do not provide the numerical table entries.
- Segmentation: Two nearby cars can be segmented as one object, while object bottoms may be assigned to the ground and an adjacent trash can to a facade.These errors are illustrated by comparison with hand segmentation in Figure 10.
- Classification: Classification is evaluated on an 80% training and 20% test split, using coarser classes and averaged precision, recall, F1, MCC, and Random-Forest OOB metrics.The authors caution that some coarse classes contain few samples.
- Classification: Geometric plus GRSD descriptors obtain the best classification results, while computing only geometric descriptors can reduce computation time for time-critical applications.The authors attribute the advantage of the smaller descriptor set to the dataset's limited sample counts and note that geometric descriptors avoid normal computation.
5 Conclusion
The authors present a large, hand-labeled urban point-cloud dataset and test an improved method for automatic segmentation and classification. Its scale is intended to support training and evaluation, including deep-learning methods.
- 140 million points span 2km across two cities, with each object manually segmented and assigned one of 50 classes.
- The dataset is intended to help train and evaluate deep-learning methods that require large numbers of points.
- The authors test a segmentation and automatic-classification method from after making robustness improvements.