Claude Code CV

A hands-on demonstration of computer vision and AI integration skills, built entirely in a single Claude Code session. The tasks — background removal, YOLO object detection, Gemini vision API integration, and comparative analysis — map to the technical stack of the Scaffluent FYP project.

Reports

• YOLO26 vs Gemini 3 Flash Comparison (PDF) — 5-metric deep dive (performance, cost, adaptability, accuracy, ease of use)
• Claude Skills Report (PDF) — Demonstrated and further hard/soft skills tied to Scaffluent's architecture

Scripts

Script	Description
net_removal.py	Net occlusion removal — L1→L2→L3 pipeline: detect safety netting via HSV colour segmentation, generate strand masks, reconstruct underlying structure via colour correction + inpainting
remove_bg.py	Background removal via flood-fill + connected component analysis
run_yolo.py	YOLO26 Nano detection on the Nyan Cat image
run_yolo_bus.py	YOLO26 Nano detection on a Madrid bus street scene
gemini_describe.py	Gemini 3 Flash image description via OpenRouter (OpenAI SDK)
gemini_structured.py	Gemini 3 Flash with enforced JSON schema output
gemini_bbox.py	Gemini 3 Flash bounding box detection + PIL visualization
render_pdf.py	Render Markdown reports to PDF via weasyprint

Images

Input

Nyan Cat pixel art	Madrid bus scene

Net Occlusion Removal

Removes safety netting from construction site images to reveal underlying scaffolding structure. Each comparison grid shows: original → region mask → strand mask → colour-corrected → strand-inpainted → combined (best).

User-sourced samples (close-up, 58–80% net coverage):

Claude-sourced samples (Wikimedia Commons, 19–25% net coverage):

Other Output

Background removed	YOLO26 detection
Gemini bbox (Nyan Cat)	Gemini bbox (bus)

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Create a .env file with your OpenRouter API key:

OPENROUTER_KEY=sk-or-v1-your-key-here

YOLO26 model weights (yolo26n.pt) are auto-downloaded on first run.

Tech Stack

• YOLO26 Nano — Local real-time object detection (Ultralytics)
• Gemini 3 Flash — Zero-shot VLM detection via OpenRouter API
• OpenCV — HSV segmentation, morphological masking, inpainting (Telea/Navier-Stokes)
• PIL/Pillow + SciPy — Image processing and visualization
• Python 3.14 — MacBook M4, bare-metal CPU inference