Concept & Overview
Lumira is an interactive computer vision web application designed to help users identify computer hardware components effortlessly. The application accepts both static image uploads and live camera feeds to deliver instantaneous on-device predictions without relying on server-side inference latency.
Engineering & Model Pipeline
- Architecture & Training: Built and trained a custom
simpleCNNclassification architecture in PyTorch, utilizing ~500 curated images per hardware class. - Cross-Platform Conversion: Converted trained PyTorch weights into ONNX format, and subsequently compiled into TensorFlow.js graph models.
- Client-Side Stability: Engineered asynchronous video frame buffer processing to ensure smooth, real-time client-side classification across both desktop browsers and mobile devices.
Collaborators & Supervision
- Team: Farrell Sudjatmiko & Reza Ali Nirwansyah
- Academic Supervision: Examined by Prof. Dr. I Ketut Eddy Purnama, S.T., M.T. (Department of Computer Engineering, ITS).