Winner - GrowthFactor Challenge @ Hacklytics 2026 (Georgia Tech, Feb 20-22)

Polaris is a scalable ML platform that quantifies parking inventory from space. Drop a pin anywhere in Atlanta, and in under 2 seconds it parses satellite imagery to deliver stall counts, confidence intervals, and a 0-100 Polaris Score for parking accessibility - across surface lots, parking garages, underground structures, and street parking.

We chose Track B (City-Wide Mapping) and mapped the entirety of Atlanta, generating a comprehensive parking heatmap across the metro area using H3 hexagonal grids.

Devpost | GitHub

• The Problem
• Our Solution
• Three Detection Approaches
• Beyond Surface Lots
• Tech Stack
• Getting Started
• API Reference
• Benchmarks
• Hackathon Context
• Team
• License

Retailers need to know one deceptively simple thing: how much parking is available near their locations?

Today, answering this question requires expensive manual surveys or proprietary datasets that are instantly stale. Satellite imagery is abundant, high-resolution, and constantly updated - but turning raw pixels into accurate parking spot counts is an unsolved data science challenge.

GrowthFactor, a company that builds software for retail and real estate decision-makers, sponsored this challenge at Hacklytics 2026. Parking availability is one of the most requested - and hardest to obtain - data points their customers ask about.

Polaris attacks the problem from three angles simultaneously, then fuses the results:

1. SegFormer-b5 - Fine-tuned on the ParkSeg12k dataset (12,617 satellite image/mask pairs across 45 US cities) for pixel-level parking lot segmentation
2. YOLOv8 / YOLO11 - Detects individual parking stalls and vehicles from satellite tiles using SAHI (Slicing Aided Hyper Inference) for small-object detection
3. Geometric Heuristics - Uses OSM polygon boundaries + ITE/NPA parking design standards (angle-specific aisle widths, OBB layout simulation) to estimate capacity from lot geometry alone - near instant and highly accurate

For structured parking invisible from above (garages, underground, street parking), we supplement with OpenStreetMap metadata, building height data, and municipal records.

The final estimate is a weighted ensemble of all methods, producing confidence intervals and the Polaris Score - a 0-100 composite metric for parking accessibility.

Fine-tuned nvidia/segformer-b5-finetuned-ade-640-640 on the ParkSeg12k dataset to produce pixel-level parking masks. The pipeline:

• Fetches 512×512 Esri satellite tiles at zoom 19
• Runs SegFormer inference with test-time augmentation (4 orientations averaged)
• Applies morphological postprocessing (close > open > filter)
• Counts stalls via connected-component analysis with area thresholds (400-12,000 px)
• Falls back to area-based estimation using real-world stall dimensions (15.5 m2/stall)

Strengths: Extremely accurate lot boundary detection, works across diverse geographies
Weaknesses: Computationally expensive, struggles to differentiate adjacent stalls

Two YOLO models run in parallel:

Both use SAHI (Slicing Aided Hyper Inference) to handle the small object sizes inherent in satellite imagery. Optimal configuration after grid search: slice_size=256×256, overlap_ratio=0.4 (MAE of 6.0 vehicles per region).

Strengths: Fast inference, directly counts individual stalls and vehicles
Weaknesses: Lower precision on densely packed lots

No ML required. Uses OSM polygon boundaries and parking engineering standards:

• OBB-based layout simulation - Fits parking rows at 45, 60, and 90 degree angles with angle-specific aisle widths (ITE/NPA standards: 7.3m for 90, 5.5m for 60, 4.0m for 45)
• Aspect-ratio detection - Narrow lots automatically switch to parallel/single-loaded layouts
• Solidity scoring - Uses convex hull ratio to penalize irregular shapes
• OSM capacity blending - When available, blends geometrically estimated values with official OSM capacity tags

Strengths: Near-instant, highly accurate for regular lots, no GPU needed
Weaknesses: Struggles with irregular shapes and lots without OSM boundaries

A key differentiator of Polaris is handling parking that satellites cannot see:

When OSM has a capacity tag, estimates are blended (60% OSM + 40% geometric) for higher accuracy.

• Python 3.13+ (with uv or pip)
• Node.js 20+ (LTS)
• Docker (for Actian VectorAI DB, optional)
• GPU recommended for SegFormer/YOLO inference (CUDA or MPS)

git clone https://github.com/V-prajit/Polaris.git
cd Polaris

# Python dependencies
pip install -r requirements.txt
# or with uv:
uv sync

# Frontend dependencies
npm install

cp .env.production .env

Edit .env and add your API keys:

GEMINI_API_KEY=your_gemini_key        # For semantic search embeddings
GOOGLE_MAPS_API_KEY=your_maps_key     # For POI enrichment (optional)
BACKEND_URL=http://localhost:8000      # API URL

docker-compose up -d

uvicorn api.app:app --host 0.0.0.0 --port 8000 --workers 2

npm run dev

Open http://localhost:3000 - you'll see the Polaris globe. Search for any Atlanta location to get a parking analysis.

For a one-click deployment on a Vultr VPS:

bash scripts/deploy_vultr.sh

This sets up systemd services for both the API and frontend, configures the firewall, and starts the VectorAI DB container.

Point-level parking analysis for a single location.

Returns: Surface lots with stall counts (SegFormer + YOLO + geometric), structured parking estimates, street parking estimates, confidence intervals, segmentation mask contours (GeoJSON), and the Polaris Score.

City-wide parking heatmap using H3 hexagonal grid.

Returns: Array of H3 hexagon cells, each with parking capacity estimates and metadata.

Semantic search over indexed parking profiles using natural language.

{
  "query": "areas with lots of garage parking near restaurants",
  "top_k": 10,
  "min_spots": 50,
  "require_garage": true
}

Returns: Top-K matching hex cells ranked by semantic similarity via Gemini embeddings and Actian VectorAI DB.

Georgia Tech, Atlanta (33.7756, -84.3963)
----------------------------------------------
  Lot #301779 (3,396 m2)
    Area heuristic:    109  [81-142]
    Edge detection:     31
    Geometric:          65  [52-82]
    YOLO spots:         84
    YOLO cars (SAHI):    7  [5-9]

Atlantic Station, Atlanta (33.757, -84.4015)
----------------------------------------------
  Lot #800491500 (6,478 m2)
    Area heuristic:    208  [156-271]
    Geometric:         122  [97-152]
    YOLO spots:         87
    YOLO cars (SAHI):    2  [1-2]

The best results come from averaging YOLO spot detection with the SegFormer-informed geometric heuristic, while other methods serve as validation cross-checks.

Best SAHI config: slice=256x256, overlap=0.4 - MAE: 6.0 vehicles/region

This project was built in 36 hours for the ParkSight challenge at Hacklytics 2026 (Georgia Tech, Feb 20-22), sponsored by GrowthFactor.

Build a pipeline that uses satellite or aerial imagery to map and count parking spots. Two tracks were offered:

• Track A (Point Query): Given a lat/long, estimate nearby parking capacity
• Track B (City-Wide Mapping): Generate a comprehensive parking map for an entire city (more ambitious, weighted favorably)

We chose Track B and delivered a full city-wide parking map of Atlanta, while also supporting point queries.

• ParkSeg12k - 12,617 satellite image/mask pairs for SegFormer training
• APKLOT - 7,000 annotated polygons for YOLO training
• VisDrone - Aerial vehicle detection dataset (pre-trained model)
• Esri World Imagery - High-res satellite tiles via Contextily
• OpenStreetMap - Parking polygons, building metadata, road networks, capacity tags
• Atlanta Zoning Districts - GeoJSON for zoning context

Sponsor Mentor: Raj - Co-founder at GrowthFactor

MIT - Copyright (c) 2026 GrowthFactor, Inc.