Summary

Netryx is an open-source, locally-hosted geolocation tool using state-of-the-art computer vision to pinpoint exact GPS coordinates from street-level images. It replicates high-end SaaS functionality, offering sub-50m accuracy without landmarks, entirely on local hardware. Ideal for researchers, journalists, and power users, it provides a privacy-preserving, cost-effective alternative to cloud services.

How It Works

Netryx employs a three-stage pipeline: 1) Global Retrieval: CosPlace extracts image descriptors, compared against a pre-computed index using cosine similarity and radius filtering for rapid candidate retrieval. 2) Geometric Verification: ALIKED (CUDA) or DISK (MPS/CPU) extract local features, matched by LightGlue, with RANSAC filtering for consistency. 3) Refinement: Optimizes matches via heading adjustments, spatial consensus, and confidence scoring. This progressive narrowing ensures high accuracy.

Quick Start & Requirements

Installation requires cloning, Python virtual environment setup, and pip install -r requirements.txt (including LightGlue from GitHub).

• OS: macOS, Linux, Windows (macOS M1+ or Linux/NVIDIA GPU recommended).
• Hardware: Min 4GB VRAM, 8GB RAM, 10GB storage (50GB+ for indexing).
• Python: 3.9+ (3.10+ rec.).
• GPU: MPS (Mac) or CUDA (NVIDIA) highly recommended; CPU fallback is slow.
• Optional: Gemini API key for AI Coarse mode. Indexing demands significant time (hours/days) and storage.

Highlighted Details

• Sub-50m accuracy without visible landmarks; fully local execution ensures data privacy.
• "Ultra Mode" enhances accuracy for challenging images (LoFTR, descriptor hopping).
• Optimized for NVIDIA CUDA and Apple Silicon MPS; incremental indexing allows process resumption.

Maintenance & Community

Led by AI researcher Sairaj Balaji, supported by Microsoft for Startups and ElevenLabs. Featured in Fast Company and Deutsche Welle. No specific community channels are detailed.

Licensing & Compatibility

Released under the MIT License. Users must comply with Google Maps/Street View ToS and local privacy laws. Prohibits use for stalking, harassment, or unauthorized surveillance.

Limitations & Caveats

Exclusively for street-level imagery; ineffective for indoor, aerial, or close-up photos. Setup and indexing are resource-intensive and time-consuming, requiring technical expertise. Users are solely responsible for legal and ethical compliance.