Idea
A synthetic lidar dataset platform providing high-fidelity annotated data for autonomous vehicle and smart city AI developers.
Research Paper
Core Innovation
This paper introduces UrbanTwin, a method to generate synthetic roadside lidar datasets using digital twins that replicate real-world geometry and vehicle movement. Unlike prior synthetic data, UrbanTwin closely matches real data distributions, improving model training and generalization. It also allows customizable scenarios to test AI models under varied conditions.
Market Size (TAM)
$2–10B TAM, $1–2B SAM; assumption: growing autonomous vehicle and smart city AI markets require extensive labeled lidar data.
Potential Customers & Pain Points
- Autonomous Vehicle Developers Needing Diverse Training Data
- Smart City Planners Requiring Accurate Traffic Models
- AI Researchers Lacking Large-Scale Annotated Lidar Datasets
Business Model
Offer datasets via subscription or licensing for research and commercial use; provide custom scenario generation services for enterprise clients.
Competitive Landscape
- Waymo Open Dataset
- nuScenes
- KITTI
Implementation Challenges
- High computational cost for digital twin simulation
- Ensuring synthetic data matches all real-world edge cases
- Adoption by industry requiring validation against real data
Validation Strategy
- Benchmark model performance trained on UrbanTwin vs real datasets
- Collaborate with autonomous vehicle companies for pilot testing
- Publish comparative studies demonstrating generalization improvements
Research Paper Overview
UrbanTwin: High-Fidelity Synthetic Replicas of Roadside Lidar Datasets
Summary
UrbanTwin datasets are realistic synthetic replicas of three public roadside lidar datasets, each containing 10K annotated frames with 3D bounding boxes, instance segmentation, tracking IDs, and semantic labels. These datasets are generated using emulated lidar sensors within precise digital twins modeled on real-world geometry, road alignment, and vehicle movement patterns. The synthetic data aligns closely with real data, enabling effective training of 3D object detection and segmentation models that generalize well to real-world scenarios. UrbanTwin enhances existing benchmarks by increasing data diversity and sample size and supports custom scenario testing through adaptable digital twins. The datasets are publicly available for research and development.