Research map / Point clouds / Point cloud classification
USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset
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 USCILab3D, a year-long outdoor robot dataset from USC campus with 10M images and 1.4M LiDAR scans, semantically labeled via GPT-4 and Grounded-SAM.
Problem
Lack of large-scale, long-term, semantically annotated outdoor 3D datasets with multi-view images and aligned LiDAR point clouds for 3D vision and robotics.
Approach
Automatically generate 3D semantic point cloud labels without manual annotation by using GPT-4 for candidate labels, Grounded-SAM for pixel masks, and projecting 2D masks onto calibrated LiDAR points, with DBSCAN cleaning.
Key results
- 3D Gaussian Splatting on USCILab3D achieves PSNR 26.02, SSIM 0.86, LPIPS 0.20, outperforming ETH3D (21.25 PSNR) and Tanks&Temples (21.20 PSNR) on the reported metrics.
- Dataset contains roughly 10M images, 1.4M semantic point cloud scans, and 267 semantic labels grouped into 12 categories, collected over 12 months on a 229-acre campus.
- Adding unrelated prompts to Grounded-SAM changed incorrect pixel labels only from 0.23% to 0.92% (1 to 10 extra prompts).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use USCILab3D rather than the comparison datasets | 3D Gaussian Splatting · Novel View Synthesis | PSNR: 26.02 vs 27.01 | Mixed |
| Use USCILab3D rather than the comparison datasets | 3D Gaussian Splatting · Novel View Synthesis | SSIM: 0.86 vs 0.87 | Mixed |
| Use USCILab3D rather than the comparison datasets | 3D Gaussian Splatting · Novel View Synthesis | LPIPS: 0.20 vs 0.27, 0.28, 0.25, 0.32 | Helped |
| Add more unrelated prompts (ablation) | Grounded-SAM · Grounded-SAM prompt robustness | Incorrect pixel labels: 0.23% -> 0.92% | Hurt |
Surprising findings
- Deep Blending achieves higher PSNR (27.01) than USCILab3D, which the authors attribute to its smaller size of 2.6K images.
Limitations
- Annotations are machine-generated with no manual verification at the time of writing; human validation was planned.
- Covers a narrow geographic scope (USC University Park campus) compared to city-scale datasets.
- Camera-LiDAR pose interpolation is imperfect; unified poses were not yet released.
Experimental setup
Datasets: USCILab3D
Benchmarks: USCILab3D ETH3D Tanks&Temples Mip-NeRF360 Deep Blending
Metrics: PSNR SSIM LPIPS
Baselines: ETH3D Tanks&Temples Mip-NeRF360 Deep Blending
Builds on: Grounded-SAM 3D Gaussian Splatting LeGO-LOAM COLMAP
Closest papers by meaning
- SynRS3D: A Synthetic Dataset for Global 3D Semantic Understanding from Monocular Remote Sensing Imagery (NeurIPS 2024 · similarity 0.88)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.87)
- SS3DM: Benchmarking Street-View Surface Reconstruction with a Synthetic 3D Mesh Dataset (NeurIPS 2024 · similarity 0.86)
- LiDAR-LLM: Exploring the Potential of Large Language Models for 3D LiDAR Understanding (AAAI 2025 · similarity 0.86)
- CymbaDiff: Structured Spatial Diffusion for Sketch-based 3D Semantic Urban Scene Generation (NeurIPS 2025 · similarity 0.86)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.86)
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.