About VisuTrack
Inspiration
We were inspired by the inefficiencies in manual visual inspection processes across industries — from factories to infrastructure sites. Traditional pixel-based methods fail under lighting or angle variations and can’t detect semantic changes like rust, cracks, or missing parts. We wanted to build a smarter, AI-powered engine that truly understands visual differences and makes inspections faster, accurate, and scalable.
What it does
VisuTrack is an AI-powered Visual Difference Engine that automatically detects, classifies, and highlights changes between time-series images. It goes beyond pixel comparison — it uses semantic AI embeddings and anomaly detection to understand what actually changed. Users can view before/after overlays, visualize heatmaps, classify severity, and export change reports — enabling real-time quality and safety monitoring.
How we built it
We built VisuTrack using a combination of FastAPI, PyTorch, and OpenCV for AI and image processing, with React, Tailwind CSS, and Recharts/D3.js powering the interactive dashboard.
Vision Transformer / ResNet embeddings for semantic change detection
Autoencoder & One-Class SVM for anomaly filtering
Edge-deployable ONNX models for offline analysis
PostgreSQL + Redis + MinIO for backend data management
Challenges we ran into
Handling lighting variations and misaligned camera angles during image comparison
Designing a semantic model that differentiates between real changes and visual noise
Processing large image datasets efficiently for near real-time inference
Building an intuitive UI to visualize complex detection outputs clearly
Accomplishments that we're proud of
Achieved 35% higher accuracy over pixel-difference methods
Reduced inspection time by up to 70%
Built a modular, edge-ready pipeline that can run offline
Delivered an interactive dashboard with overlays, severity maps, and exportable reports
What we learned
Deep understanding of semantic embeddings, temporal vision, and anomaly detection
How to balance heavy AI processing with edge performance constraints
Importance of user feedback loops in improving AI models over time
Practical DevOps for deploying AI on lightweight, edge environments
What's next for VisuTrack
Integration with IoT edge cameras and drone-based inspections
Expansion into 3D change detection using depth maps
Federated learning for privacy-first model updates
AR-assisted dashboards for real-time overlay in physical environments
Log in or sign up for Devpost to join the conversation.