Comparative Analysis: AVIS vs. RoadX
This document provides a comprehensive, feature-by-feature comparison between AVIS (our system) and RoadX (the competitor project), analyzing their architectural choices, strengths, weaknesses, and model implementations.
1. High-Level Architecture & Scalability
| Feature |
AVIS (Gridlock) |
RoadX |
Winner & Why |
| System Design |
Modular Monolith (FastAPI + React) |
Single-App Monolith (Flask) |
AVIS. A decoupled frontend and backend is modern and scalable. RoadX relies on Flask HTML templates, which limits UI reactivity. |
| Asynchronous Processing |
Dedicated Redis Queue + Background Workers |
Python Background Threads |
AVIS. Using Redis ensures heavy CV tasks don't block the API. RoadX's threading on a single Flask app will bottleneck and drop frames under heavy load. |
| Data & Storage |
PostgreSQL + MinIO (S3-compatible) |
SQLite + Local File System |
AVIS. AVIS is built for cloud/production scale. RoadX is limited to a single server's disk space. |
2. Core Machine Learning & Logic
| Feature |
AVIS (Gridlock) |
RoadX |
Winner & Why |
| Verification & Accuracy |
Hybrid: CV + VLM (Gemini Flash) |
Pure CV (YOLOv8 + ByteTrack) |
AVIS (for accuracy). RoadX relies entirely on geometric CV, which has high false-positive rates for edge cases. AVIS uses Gemini as a final judge, ensuring court-ready evidence. |
| Video & RTSP Support |
❌ Single Image Only |
✅ RTSP Streams & Video |
RoadX. By using ByteTrack (assigning IDs to vehicles across frames), RoadX can easily detect temporal violations (Wrong-Way, Speeding) without needing a VLM. |
| ALPR (Number Plates) |
fast-alpr (Generic) |
Custom Plate.pt + EasyOCR |
RoadX. RoadX trained a custom YOLO model specifically to find Indian plates, then runs OCR on the crop. This is vastly superior to generic out-of-the-box ALPR tools. |
| Decision Logic |
Deterministic Evidence Graph |
Imperative Python logic |
AVIS. Building a node-based "Evidence Graph" is much more explainable in a court of law than nested if/else Python statements. |
3. Features & End-User Experience
| Feature |
AVIS (Gridlock) |
RoadX |
Winner & Why |
| User Interface |
Hardware-accelerated, dynamic React UI |
Basic HTML/Bootstrap Flask templates |
AVIS. The observability trace logs, empty-state hero, and modern theming in AVIS provide a vastly superior enterprise feel. |
| End-to-End Enforcement |
Generates database records |
PDF Challans + Email/WhatsApp |
RoadX. RoadX completes the enforcement loop by actually generating a legal PDF and notifying the citizen. |
| Access Control |
Single unified dashboard |
Admin vs. Citizen Portal |
RoadX. Separating the dashboard into a secure Admin police view and a public citizen ticket-checking view is highly practical. |
Summary of Pros & Cons
AVIS (Our System)
- Pros: Exceptionally scalable architecture (Redis, MinIO, Postgres). Highly accurate due to VLM (Gemini) fallback logic preventing hallucinations. Beautiful, observable, enterprise-grade React frontend.
- Cons: Lacks temporal tracking (video processing). Relies on generic ALPR which struggles with dirty/angled Indian plates. Incomplete enforcement loop (no PDFs or emails).
RoadX (Competitor)
- Pros: Handles live RTSP video natively using ByteTrack. Excellent Indian license plate detection via custom YOLO weights. Generates real-world PDF challans and emails.
- Cons: Monolithic Flask app will choke on high load. Relies entirely on CV rules, meaning it will likely generate false-positive challans that a human/VLM would easily reject.
How RoadX Handled Video (The "Secret")
You asked how they handle video: They use an algorithm called ByteTrack.
Instead of just finding a car, ByteTrack assigns an ID to a car in Frame 1 (e.g., Car #45). In Frame 2, it finds Car #45 again and draws a line between its old and new position. By tracking this trajectory over 20-30 frames, they can easily calculate if the car is moving in the wrong direction (Wrong-Way) or how fast it is moving, without needing complex AI reasoning.
