csg
Résumé ↗
Menu
← All projects

Computer vision / Prototype

Fridge Vision

A fridge photo. A dinner idea.

Detect ingredients in a photo, read package labels when needed, and turn structured detections into recipe suggestions.

YOLO11 · FastAPI · OpenCV · React

View repository ↗

01 / THE PROBLEM

What needed solving.

A photo of a fridge is unstructured input. Recipe search needs ingredient names. This project connects a computer-vision model to a complete upload, detection and recommendation flow, with a local matching path and optional language-model generation.

02 / FOLLOW THE WORKFLOW

See how it fits together.

A simplified example of the documented flow. Sample outputs are authored illustrations, not live model predictions.

Fridge VisionIllustrative workflow

Illustrative image contents

Tomato · Egg · Cheese

Upload a fridge image to the application.

Example detection labels

[ Tomato ] [ Egg ] [ Cheese ]

YOLO identifies ingredients; OCR can read labels.

Structured ingredient list

tomato + egg + cheese

Pass detected ingredients to the recipe matcher.

Illustrative recipe recommendation

Tomato & cheese omelette

Browse a matching recipe suggestion.

03 / ENGINEERING

The decisions behind it.

Keep detection behind an API

The React interface sends the image to FastAPI. Ultralytics YOLO runs detection, while OpenCV and optional EasyOCR support image and label processing.

Use structured results for recommendations

Detected labels feed a local keyword matcher. Optional Ollama generation is a separate path, so the basic recipe workflow does not require a language-model service.

Treat quantities as approximate

Bounding boxes provide simple quantity estimates. They do not establish food weight, volume or nutritional content, especially under occlusion or inconsistent camera scale.

04 / EVIDENCE & LIMITS

Claims with context.

67Configured food classes
2Recipe paths: local matching or optional LLM

The README specifies YOLOv11-L, 800-pixel inference, confidence 0.25 and IoU 0.5. These are configuration values, not accuracy measurements. The repository includes a React upload experience and separate ingredient, recipe and combined API endpoints.

Current limits & next validation

No held-out mAP, per-class precision/recall or end-to-end latency benchmark was established in the reviewed documentation. Lighting, occlusion and packaging need testing. Recipe suggestions and approximate quantities do not establish dietary suitability or reduced food waste.

05 / READ THE SOURCE

Go deeper.

Based on public repository documentation reviewed on 7 November 2026. Results are attributed to the project’s recorded evaluations; they were not rerun for this portfolio.

NEXT PROJECTVoice-Based RAG

TOOLS BEHIND THE WORK

Explore the stack. Drag the ribbon or select a tool.