The second generation of the Live Game Defender (LGD) playing-card detector: gen 1 fine-tuned on the first day of real casino-table video from our proof-of-concept rig. This closed most of the synthetic→real gap that held gen 1 back on real footage.

Superseded by lgd-cards-gen3 (day-2 video, holdout recall 0.85 — the current served model). Gen 2 is published for provenance/reproducibility.

Generations

Gen	Repo	Trained on	Frozen real-video holdout recall	Status
1	lgd-cards-gen1	Roboflow `ow27d` v4 dataset	— (dataset-val only)	superseded
2	lgd-cards-gen2 (this)	+ day-1 PoC table video	0.68	superseded
3	lgd-cards-gen3	+ day-2 PoC table video	0.85	✅ current

Chip detectors: lgd-chips-gen1 · lgd-chips-gen2.

Classes (52)

10C 10D 10H 10S 2C 2D 2H 2S 3C 3D 3H 3S 4C 4D 4H 4S 5C 5D 5H 5S
6C 6D 6H 6S 7C 7D 7H 7S 8C 8D 8H 8S 9C 9D 9H 9S
AC AD AH AS JC JD JH JS KC KD KH KS QC QD QH QS

C=Clubs, D=Diamonds, H=Hearts, S=Spades. Full order is in model.classes.json.

Files

model.onnx — ONNX export (run with onnxruntime).
model.classes.json — ordered class-name sidecar (index → card code).
metrics.json — training config + full per-label holdout evaluation.

Training

Base: Ultralytics yolo11s.pt (COCO-pretrained), 40 epochs (early-stop patience 15), imgsz=640.
Data: Roboflow ow27d base set + day-1 PoC table recordings, auto-labeled (pip boxes proposed by the gen-1/ow27d detector, verified/named by an LLM over a closed 52-code vocabulary), materialized as serve-matching 2×2 tiles (3,112 train / 445 val tiles). Labeling cost $3.41.
Hardware: NVIDIA RTX 3060 (12 GB).

Metrics — frozen real-video holdout

Scored on a frozen 117-frame holdout of our own PoC recordings (never trained on), serve-style auto-grid tiling, threshold 50, code-aware:

Recall	Precision proxy
0.68	0.604

⚠️ Not casino accuracy (rule of the project). The holdout is our own proof-of-concept footage; ground-truth labels are LLM-verified, not fully human-verified. This demonstrates the detector mechanism and the generation-over-generation improvement, not a real-world full-deck accuracy claim. Dataset-internal mAP@50 (0.81) is in metrics.json and is likewise not real-world.

Usage

python

import json, onnxruntime as ort
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
names = json.load(open("model.classes.json"))   # index -> "AS", "10H", ...
# 640x640 letterboxed input; this model detects corner PIPS — cluster same-code pips into one card.

License & provenance

AGPL-3.0, a fine-tune of Ultralytics YOLO11 (yolo11s.pt, AGPL-3.0) — these weights inherit AGPL-3.0 and are not an original work of ours. Networked deployment triggers AGPL §13 (offer the Corresponding Source). onnxruntime (MIT) keeps the inference code AGPL-free; the weights stay AGPL.

sroot/lgd-cards-gen2

LGD Cards — Gen 2 (day-1 video) · YOLO11s playing-card detector, 52 classes