dhvazquez/mtg_card_identification_embeddings

MTG Card Identification Model

Goal

Given a cropped MTG card image, produce an embedding for similarity-based identification against a reference gallery.

Recommended Model: 20260521-044732/onnx/mtg_embed_fp16.onnx

Resources

Pose Model: https://huggingface.co/dhvazquez/mtg_card_pose_estimation
Pose Train: https://github.com/diegovazquez/mtg_train_card_pose_estimation

Embedding Model:https://huggingface.co/dhvazquez/mtg_card_identification_embeddings
Embedding Train: https://github.com/diegovazquez/mtg_train_card_identification_embeddings

Dataset: https://huggingface.co/datasets/dhvazquez/mtg_synthetic_large_dataset
Dataset Generator: https://github.com/diegovazquez/mtg_synthetic_dataset_generator

Pipeline

Image → Pose ONNX (warp) → Embedder ONNX → 256-d vector → similarity search

Embedding Model

Attribute	Value
Architecture	EfficientNet-Lite0 + SubCenterArcFace
Input	448×320 px (H×W)
Output	256-dimensional L2-normalized vector
Training classes	82,115 unique cards
Training	24 epochs, 2× GPU, batch 128

Metrics (Validation Set)

Metric	Value
Recall@1	94.1%
Recall@5	94.6%
mAP	94.3%

Exported Variants (ONNX)

Variant	Size	Use case
fp32	14.1 MB	Reference / maximum precision
fp16	7.0 MB	GPU / edge production
int8	3.7 MB	CPU / mobile production