Overview
This section documents our statewide, AI-driven sidewalk detection pipeline that transforms high-resolution aerial imagery into accurate, vectorized sidewalk networks. The system spans the full machine learning lifecycle — from dataset curation and model training, to scalable inference and statewide deployment.Introduction
Project overview and system architecture
GAIS Lakehouse
GIS data formats and the open lakehouse stack
Models
DeepLabV3+
Our primary semantic segmentation model
Mask R-CNN
Alternative Detectron2 model architecture
SAM
Segment Anything Model exploration
Training
Training Pipelines
Distributed training with Ray and ClearML
Dataset Processing
Data cleaning, augmentation, and streaming
Metrics Report
Training results and model performance
Inference
Pipeline Architecture
Ray + Triton distributed inference system
NJ Statewide Results
Production deployment and performance analysis
Key Achievements
- 9,202 tiles processed in a single statewide run
- 4.8 million chips with ~319 billion pixels analyzed
- ~17 hours for complete state coverage on single GPU
- 99.66% success rate with robust fault handling
- 96% pixel accuracy and 0.77 mean IoU on validation
Technology Stack
- Model: DeepLabV3+ with ResNet-103 backbone
- Training: Detectron2 with Hugging Face streaming
- Inference: Ray actors + NVIDIA Triton Inference Server
- Output: GeoJSON centerlines for transportation planning