Research map / Point clouds / Point cloud classification
LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces LiDAR-Net, a real-scanned indoor point cloud dataset (~3.6B annotated raw points, 30,000 m², 24 semantic classes) with semantic, instance and detection labels, and benchmarks standard 3D methods on it.
Problem
Existing indoor 3D datasets (S3DIS, ScanNet, ScanNet++) are built from RGB-D captures or annotated reconstructed meshes, so they lack the non-uniform distribution (scanning holes/lines) and scanning anomalies (reflection noise, ghosts) of real LiDAR point clouds, which hurts deep models on real scan
Approach
A large indoor dataset built from raw points captured by a Leica BLK2GO LiDAR scanner, with point-level semantic and instance labels directly on raw points, including annotated reflection noise and ghost anomalies, and benchmarks showing models trained on it transfer better to real-scanned rooms.
Key results
- On LiDAR-Net, KPConv reaches 44.29 mIoU and RandLA-Net 32.60 mIoU; GroupFree3D reaches 46.80 mAP@0.25 and 18.86 mAP@0.5 for detection.
- Models trained on LiDAR-Net and tested on 5 real-scanned rooms reach 39.75 mIoU, versus 35.58 (S3DIS), 33.18 (ScanNet) and 23.60 (ScanNet++) for semantic segmentation.
- Per-class IoU varies widely across methods; for example, RandLA-Net scores 42.02 on tree versus 7.94 for KPConv, while KPConv scores 58.02 on sofa versus 0.00 for RandLA-Net.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train RandLA-Net on LiDAR-Net instead of S3DIS, ScanNet, or ScanNet++ | RandLA-Net · Real-scanned data (5 rooms) | mIoU: 35.58 -> 39.75; 33.18 -> 39.75; 23.60 -> 39.75 | Helped |
| Train ASIS on LiDAR-Net instead of S3DIS, ScanNet, or ScanNet++ | ASIS · Real-scanned data (5 rooms) | mPre: 15.72 -> 32.92; 15.88 -> 32.92; 26.48 -> 32.92 | Helped |
| Train ASIS on LiDAR-Net instead of S3DIS, ScanNet, or ScanNet++ | ASIS · Real-scanned data (5 rooms) | mRec: 7.84 -> 20.39; 14.44 -> 20.39; 4.68 -> 20.39 | Helped |
| Train GroupFree3D on LiDAR-Net instead of S3DIS, ScanNet, or ScanNet++ | GroupFree3D · Real-scanned data (5 rooms) | mAP@0.25: 7.93 -> 32.96; 20.91 -> 32.96; 13.31 -> 32.96 | Helped |
| Train GroupFree3D on LiDAR-Net instead of S3DIS, ScanNet, or ScanNet++ | GroupFree3D · Real-scanned data (5 rooms) | mAP@0.5: 0.28 -> 9.22; 1.21 -> 9.22; 1.67 -> 9.22 | Helped |
Surprising findings
- No single benchmark method consistently outperforms the others across semantic classes.
- RandLA-Net gets 0.00 IoU on several classes (wardrobe, bed, sofa, reflection noise) despite reasonable overall performance.
Limitations
- Overall segmentation and detection performance remains unsatisfactory on the dataset.
- Test-set labels are hidden and evaluation is via an online platform, so results depend on that server.
- The paper explicitly does not propose a new method for its identified challenges; the challenges are discussed rather than solved.
Experimental setup
Datasets: LiDAR-Net S3DIS ScanNet ScanNet++
Benchmarks: LiDAR-Net
Metrics: mIoU mPre mRec mAP@0.25 mAP@0.5
Baselines: PointNet PointNet++ KPConv RandLA-Net ASIS 3D-BoNet VoteNet GroupFree3D
Builds on: KPConv RandLA-Net VoteNet GroupFree3D ASIS
Closest papers by meaning
- Not All Points Are Equal: Learning Highly Efficient Point-Based Detectors for 3D LiDAR Point Clouds (CVPR 2022 · similarity 0.89)
- Large-Scale Point Cloud Semantic Segmentation With Superpoint Graphs (CVPR 2018 · similarity 0.88)
- iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point Clouds (AAAI 2024 · similarity 0.88)
- LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.88)
- CurveCloudNet: Processing Point Clouds with 1D Structure (CVPR 2024 · similarity 0.87)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.