PictureCram

Self-contained localhost webapp for curating computer vision training datasets.

• Define categories,
• Capture frames from network cameras
• Run teacher models for auto-labeling
• Review and edit detection-boxes and categories

Future Plans

• Execute Training Runs
• View Live CV Inference
• Generate and Alert on Real-Time Telemetry
• Video 🚀
• Discord Integration

Quick Start

git clone <this-repo>
cd pictureCram
pip install -e .
python -m picturecram --project ~/my-dataset
# Open http://localhost:8777

Design Principles

• Self-contained — single pip install, runs on localhost, zero cloud dependencies
• Flat files — sidecar JSONs are the source of truth; SQLite is a rebuildable index
• Framework-agnostic — register YOLO, DINO, or custom script models
• Distillation-first — teacher models generate suggestions, you review & correct, then export to train smaller models
• Ease-of-use — single click edits. Make best use of human annotation time

Requirements

• Python 3.10+
• OpenCV (for stream ingestion)
• Pillow (for annotation rendering)
• Optional: ultralytics, timm, or torch (for running teacher models — only needed for the frameworks you use)

Project Status

🚧 Under active development.