TCG Scanner

Note

Work in progress

TCG Scanner is an end-to-end computer vision pipeline for real-time trading card detection and OCR, built with production-grade MLOps/Data Engineering practices.

A YOLO segmentation model localizes cards in the camera feed, applies inverse perspective warping for precise extraction, and feeds the result into a GPU-accelerated PaddleOCR service for text recognition — all running in a Dockerized microservice with a REST API.

Tech Stack: Python · PyTorch · YOLOv11 · OpenCV · PaddleOCR · Docker · Google Cloud Storage · BigQuery · CUDA · ONNX Runtime

Requirements

• Docker + Docker Compose
• NVIDIA GPU + nvidia-container-toolkit

sudo apt install nvidia-container-toolkit
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

Usage

docker compose up -d
python -m client.inference.main

The first build takes a few minutes — PaddleOCR models are downloaded and baked into the image. Subsequent starts are fast.

docker compose logs -f
docker compose down

Related

• tcgscanner-ml — Training pipeline, synthetic data generation, and model weights