The current generation of the Live Game Defender (LGD) playing-card detector, and the model LGD serves in production. Gen 2 further fine-tuned on a second day of real casino-table video (plus a human-QA correction pass), lifting frozen-holdout recall to 0.85. Locates every playing card in a frame and names it by rank + suit (52 classes); an empty frame yields no detection (built-in "is there a card?" gate).

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1	lgd-cards-gen1	Roboflow `ow27d` v4 dataset	— (dataset-val only)	superseded
2	lgd-cards-gen2	+ day-1 PoC table video	0.68	superseded
3	lgd-cards-gen3 (this)	+ day-2 PoC table video	0.85	✅ current — the production/served model
4	lgd-cards-gen4	+ day-3 deck-spread video	0.786 ¹	spread specialist — activated 2026-07-20, reverted same day (gameplay regressions); not production-approved

¹ gen4's 0.786 is on an expanded 318-frame holdout where gen3 scores 0.750 — compare only within one suite. Naming unified 2026-07-20: genN means the same model on HF, on disk and in every doc.

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); this is the exact file LGD serves.
best.pt — Ultralytics PyTorch weights (source for the ONNX export; use for further fine-tuning).
model.classes.json — ordered class-name sidecar (index → card code).
data.yaml — class list / dataset config.
train-args.yaml — the exact Ultralytics training arguments for this run.
metrics.json — frozen-holdout verdict vs. the previous generation.

Training

Base: Ultralytics yolo11s.pt (COCO-pretrained), imgsz=640 (resumed once from epoch 15).
Data: Roboflow ow27d base + day-1 PoC video + day-2 PoC video, auto-labeled (detector proposes pip boxes; an LLM verifies/names over a closed 52-code vocabulary) and materialized as serve-matching tiles — 7,526 train / 1,103 val tiles, with 14 human-QA label corrections. Labeling cost $8.13.
Hardware: NVIDIA RTX 3060 (12 GB).

Metrics — frozen real-video holdout

Scored on a frozen 225-frame holdout of our own PoC recordings (never trained on), serve-style auto-grid tiling, threshold 50, code-aware. Both generations scored in one run:

Model	Recall	Precision proxy
gen 3 (this)	0.847	0.771
gen 2	0.745	0.759

Generation-over-generation on the frozen holdout: stock 0.539 → gen1 0.680 → gen3 0.847.

⚠️ Not casino accuracy (rule of the project). The holdout is our own proof-of-concept footage with LLM-verified (not fully human-verified) ground truth. These numbers show the detector mechanism and its improvement across generations — not a validated real-world full-deck accuracy claim on a live casino floor.

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. The detector boxes corner PIPS (~2 per card): cluster same-code pips
# within a few box-diagonals into ONE card-level detection before consuming results.

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-gen3

LGD Cards — Gen 3 (day-2 video) · YOLO11s playing-card detector, 52 classes ✅ CURRENT