ThermlVision

Overview

ThermlVision is a computer vision system that runs YOLO-based human detection on thermal imagery and video feeds, logging detections to a Supabase backend with timestamps and confidence scores.

Why I built it

Built as a contract project for a security-adjacent use case. The client needed something that could process thermal camera output without cloud latency.

Technical highlights

• –Custom-tuned YOLO model weights for thermal (non-RGB) input
• –FastAPI backend with async Supabase writes for detection logs
• –Confidence threshold tuning to reduce false positives in low-contrast scenes
• –Full pytest suite covering detection pipeline and API endpoints

Hardest engineering problem

Thermal imagery has very different contrast characteristics from RGB. The default YOLO weights performed poorly, so I built a small labeled dataset and fine-tuned for the thermal domain.

What I learned

Domain adaptation for computer vision is often more impactful than model architecture changes. Good labeled data beats model complexity.

Stack

Links