🛞
Scan Tire AI — Roadmap
Automated tire inspection system based on computer vision and deep learning. An independent, solo-founder deep-tech initiative focused on practical AI inspection.
60%
Wear Accuracy
4
ONNX Models
Phase 1
Current Stage
Solo
Team Size
Current Project Status (March 2026)
ACTIVE🎯 Strategic Decision: WEAR GRADING FIRST
OCR and sidewall analysis (e.g., reading "205/55 R16") have been postponed to Phase 2.
The immediate MVP focus is exclusively on Tread Wear Analysis.
Answering "How much life is left in this tire?" provides faster, more tangible value to users than reading the size.
Infrastructure & Public Presence
scantire.com live (HTTPS), dark tech landing page with real tire mockups, GitHub Pages hosting.
Labeling Tool (PyQt6)
Hybrid architecture (C++ preprocessing + Python ML) for dataset preparation.
ONNX Models Exported
Tread Detector, TWI Detector, Depth Model, and Wear Classifier successfully packed as single-file ONNX models.
Next.js Web Inference
Integrating onnxruntime-web. Fixing depth preprocessing (absolute mm scale) and building WearAnnotator UI.
Wear Classifier Accuracy
Currently at ~60% with ResNet18 on a small dataset (~300 photos). Needs more data to reach production quality.
Architecture & Tech Stack
⚠️ Strict Preprocessing Rules
Tread/TWI/Wear models require ImageNet normalization.
Depth model requires ONLY /255.0 normalization.
Depth scale is absolute (output * 10.0 = mm), NOT min/max normalized per frame.
| Component | Technology | Status |
|---|---|---|
| Web Application | Next.js 14+ (App Router), TypeScript, Tailwind | In Progress |
| Inference Engine | onnxruntime-web (Client-side) | In Progress |
| ML Models | PyTorch → ONNX (ResNet18 based) | Ready |
| Labeling Tool | PyQt6 + Python ML + C++ Preprocessing | Ready |
🧠 ONNX Models Pipeline
🖥️ Current Interface Implementation
The interface has evolved from a concept landing page into a functional AI-driven inspection experience. It features:
- Modern tech-style landing screen with dark cyber/matrix aesthetics
- Drag & drop, file selection, and camera photo upload options
- AR-style scanning mode with corner markers and animated neonic scanline
- Visual result display with depth overlay and color-coded wear indicators
- Statistics panel showing wear class, confidence, and depth measurements
- Manual review panel for user correction (UI stage, pending backend implementation)
Inference Flow
1. Full Image → tread_detector_web.onnx → Crop ROI 2. ROI → depth_model_web.onnx → Depth Map (0..1) → * 10.0 = Depth in mm 3. ROI → wear_classifier_web.onnx → Class (New/Good/Medium/Worn/Critical) 4. ROI → twi_detector_web.onnx → TWI Points (Optional) 5. Render Overlay → Canvas API (Color-coded by absolute depth)
Phase 1: Wear MVP (Current)
IN PROGRESS🎯 Goal: Functional Web Inference & Demonstration
Deliver a clear, practical demonstration of wear analysis.
User uploads a photo → gets a color-coded depth overlay + wear class + depth in mm.
Fix Next.js Inference Pipeline
Correct depth preprocessing (/255 only) and postprocessing (absolute mm scale).
Build WearAnnotator UI
Create component with photo preview, depth overlay, statistics (avg/min/max depth), and manual correction selector.
Improve Wear Dataset
Expand the current ~300 photo dataset to improve the 60% baseline accuracy.
Formalize Wear Methodology
Document the exact criteria for New (>=7mm), Good (5-6.9mm), Medium (3-4.9mm), Worn (1.6-2.9mm), and Critical (<1.6mm).
Phase 2: Sidewall / OCR
POSTPONED (6-12 mo)⚠️ Dependent on Phase 1 Stability
OCR requires a complex chain: block detection (YOLOX), segmentation, and character recognition (MLP).
This will only be tackled after the Wear MVP is stable and providing value.
Resume YOLOX Training
Fine-tune YOLOX for sidewall block detection (width, profile, diameter, load index).
Expand OCR Dataset
Grow the current ~500 symbol dataset to cover all edge cases and lighting conditions.
Full Reading Pipeline
Combine YOLOX + OCR to read full markings (e.g., "205/55 R16").
Phase 3: Inspection Platform
FUTUREB2B API Integration
Provide API endpoints for tire shops, fleet managers, and automotive marketplaces.
White-label Solutions
Allow businesses to use the inspection tech under their own branding.
Comprehensive Web Tool
A full dashboard for service centers to track and generate reports on tire inspections.
Current Interface: Terminator Vision UI
ACTIVE🎯 Current User Experience
The current web interface demonstrates a working AI inspection flow with a tactical "Terminator Vision" style. The product presentation has evolved from a concept landing page into a functional AI-driven inspection experience.
Modern Tech-Style Landing Screen
Dark cyber/matrix aesthetics with neonic green accents and tech-inspired visual elements.
Multi-Option Upload Interface
Support for drag & drop, file selection from gallery, and direct camera photo upload.
AR-Style Scanning Animation
During analysis, interface switches to AR-style mode with corner markers and animated neonic scanline.
Visual Result Display
Shows original photo, detected tread area with bounding box, and depth overlay with color-coded wear indicators.
Statistics Panel
Displays wear class, confidence percentage, average/min/max depth, and processing time.
Manual Review Panel
UI for manual wear class selection and correction is implemented, but backend storage is planned for a later stage.
🎨 Color-Coded Wear Visualization
| Wear Class | Depth Range | Color | Visual Indicator |
|---|---|---|---|
| New | ≥ 7.0 mm | #22c55e | Green overlay |
| Good | 5.0-6.9 mm | #84cc16 | Lime overlay |
| Medium | 3.0-4.9 mm | #eab308 | Yellow overlay |
| Worn | 1.6-2.9 mm | #f97316 | Orange overlay |
| Critical | < 1.6 mm | #ef4444 | Red overlay |
Founder's Note
"I'm Silver Kod, a 48-year-old tire shop seller from Russia. Scan Tire AI is my solo-founder, indie project built out of practical necessity and enthusiasm for computer vision.
This isn't a heavily funded startup with a massive ML team. It's a realistic, step-by-step engineering effort to solve a real problem I see every day: automating tire condition assessment. We start with tread wear because it answers the most critical question for drivers and shops alike. The rest will follow."
"I'm particularly proud of the 'Terminator Vision' interface we've developed. It gives users a tactical, high-tech experience while delivering practical results. The visual feedback system with color-coded wear indicators makes the AI's assessment immediately understandable."
This isn't a heavily funded startup with a massive ML team. It's a realistic, step-by-step engineering effort to solve a real problem I see every day: automating tire condition assessment. We start with tread wear because it answers the most critical question for drivers and shops alike. The rest will follow."
"I'm particularly proud of the 'Terminator Vision' interface we've developed. It gives users a tactical, high-tech experience while delivering practical results. The visual feedback system with color-coded wear indicators makes the AI's assessment immediately understandable."