TCGscanner-detector

TCGscanner Card Detector

This repository contains the current ONNX card-boundary detector used by the TCGscanner prototype.

The model is a single-class YOLO detector. Its task is to localize the physical trading card in a camera frame or photograph. Card identification is handled separately by SigLIP 2 visual embeddings and LanceDB vector search in the application repository.

Files

• riftbound_regions.onnx: exported ONNX detector expected at models/riftbound_regions.onnx by the scanner.

Current Artifact

• Size: 11.70 MB
• SHA256: 8566d3c8556183c780eab0937f65d9862bdfc57697bfd3c33135218f28230f41
• Class labels: card
• Default confidence threshold in the app: 0.35

Training Summary

The detector was trained on a universal TCG detection dataset that combines localized card examples from multiple trading card domains. The objective is to learn generic card geometry rather than the visual identity of a specific game.

The selected hybrid experiment used corners, polygons, and isolated full-card samples. The June 27, 2026 audit run reported:

The selected hybrid run was stopped manually during epoch 42 after the validation curve had stabilized for the scanner use case. Its best validation checkpoint was epoch 40 with mAP50=0.9942 and mAP50-95=0.9628.

Usage

uv run python scripts/download_detector.py

The scanner loads the downloaded model from:

models/riftbound_regions.onnx

Limitations

• This detector only localizes the card boundary.
• It does not identify the card.
• The current dataset still needs more real-world Riftbound photographs.
• Pricing and collection features are outside this model repository.