CSPC-Dataset: New LiDAR Point Cloud Dataset and Benchmark for Large-Scale Scene Semantic Segmentation

Guofeng Tong, Yong Li, Dong Chen, Qi Sun, Wei Cao, Guiqiu XiangIEEE Access 2020 · DOI

Introduction

Semantic segmentation—the process of assigning a category label to every point in a 3D scan—is essential for autonomous driving, urban planning, and digital twin creation. However, the field faces a significant bottleneck: the lack of diverse, large-scale, and colorized outdoor datasets captured from a pedestrian perspective. Existing benchmarks often suffer from indoor bias, occlusions from vehicle-mounted systems, or a lack of semantic context.

Comparison of existing 3D datasets (a-e) with the proposed CSPC-Dataset (f).

The authors address these gaps by introducing the CSPC-Dataset, a massive repository containing 68 million points across five distinct urban and rural environments. By utilizing a wearable mapping system, the dataset provides a "complete" view of the world, including narrow alleys and building interiors often missed by car-based LiDAR.

The CSPC-Dataset

The CSPC-Dataset covers a wide variety of scenes, including simple streets, busy urban centers (captured during both day and night), university campuses, and rural landscapes.

A sample scene from the CSPC-Dataset, demonstrating the dense, colorized point cloud captured from a pedestrian perspective.

Key features of the dataset include:

  • Pedestrian Perspective: Captured via a custom wearable backpack mapping robot, reducing occlusions common in vehicle-based scans and allowing for data collection in pedestrian-only zones.
  • Rich Context: Dense, colorized 3D data (XYZ + RGB) across 6 semantic classes: Ground, Building, Car, Bridge, Vegetation, and Pole.
  • Scale: 68 million points providing a rigorous testing ground for large-scale segmentation algorithms.

Methodology

The data was collected using a custom-built Backpack Mobile Mapping System. This system integrates two 16-beam LiDAR scanners—one mounted horizontally and another tilted at 45° to maximize vertical field-of-view—with a panoramic camera for point cloud colorization. An onboard IMU and controller facilitate real-time SLAM, allowing for mapping in areas inaccessible to vehicles.

To label the resulting 68 million points, the researchers employed a semi-automated pipeline:

  1. Ground Extraction: Automated terrain separation using a "Cloth Simulation Filtering" algorithm.
  2. Split and Merge: A manual strategy to handle complex objects by partitioning large scenes into manageable patches for labeling and then re-integrating them.
  3. Refinement: A rigorous double-check process to eliminate noise and ensure ground-truth accuracy.

Benchmark Results

The authors evaluated seven baseline methods, including traditional Machine Learning (ML) and modern Deep Learning (DL) architectures like PointNet++, SnapNet, and Label-3D CNN.

  • Deep Learning Dominance: PointNet++ emerged as the top performer, achieving an Overall Accuracy (OA) of up to 96.6% in specific scenes. Its ability to learn high-level features directly from raw point sets proved superior to traditional methods as training data increased.
  • Class-Specific Insights: While "Ground" and "Building" segmentation reached high Intersection over Union (IoU) scores (97.8% and 93.7%, respectively), "Poles" remained a significant challenge with a top IoU of only 36.0%. This is attributed to class imbalance and the geometric similarity of poles to other vertical structures.
  • The Value of Color: Providing RGB values alongside XYZ coordinates significantly helped algorithms distinguish between objects with similar geometries but different appearances, such as grey poles versus thin tree trunks.

Conclusion

The CSPC-Dataset provides a vital new benchmark for the 3D vision community. By shifting the focus to pedestrian-level, large-scale outdoor scenes, it highlights both the progress and the remaining hurdles—such as long-tail recognition for rare objects and algorithmic generalization across diverse topographies. Future work will focus on improving model performance in "messy" real-world environments without requiring massive retraining for every new scene.

Summary updated

BibTeX
@article{tong2020cspc,
  author = {Guofeng Tong and Yong Li and Dong Chen and Qi Sun and Wei Cao and Guiqiu Xiang},
  title = {{CSPC-Dataset: New LiDAR Point Cloud Dataset and Benchmark for Large-Scale Scene Semantic Segmentation}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2992612},
}
Figure