sroot/lgd-chips-gen2

LGD Chips — Gen 2 · YOLO11s casino chip-colour detector (5 colours, dual-felt) ✅ CURRENT

The current generation of the Live Game Defender (LGD) casino-chip detector, and the model LGD serves in production. A YOLO11s object detector that finds gaming chips on the felt and classifies each by colour. Gen 2 is a dual-felt model — trained on a blend of Ultimate Texas Hold'em and blackjack footage so it works on both table types — and uses a venue-matched palette. Colours only; colour→denomination is per-casino configuration downstream.

The palette is the actual venue's chips: chip_black chip_green chip_pink chip_purple chip_yellow — no white, no red (the light "100" chip is the black denomination; felt red is only printed markers and card suit-pips).

Generations

Gen	Repo	Palette	Frozen-gate verdict	Status
1	lgd-chips-gen1	black / white / green / red / pink	blackjack gate: 0.91 / 0.77	superseded
2	lgd-chips-gen2 (this)	black / green / pink / purple / yellow	UTH gate: 0.80 / 0.85 / colour 0.97	✅ current

Card detectors: lgd-cards-gen1 · lgd-cards-gen2 · lgd-cards-gen3.

Classes (5, in model.classes.json order)

chip_black  chip_green  chip_pink  chip_purple  chip_yellow

Files

• model.onnx — ONNX export (run with onnxruntime); this is the exact file LGD serves.
• model.classes.json — ordered class-name sidecar (index → colour).
• metrics.json — training config + evaluation.

Training

• Base: Ultralytics yolo11s.pt (COCO-pretrained), 60 epochs, imgsz=640.
• Data: a blend of the UTH chip corpus and the gen-1 blackjack data (point-labeled by vision agents), making it a dual-felt model.
• Hardware: NVIDIA RTX 3060 (12 GB).

Metrics — frozen gates (never trained on)

Gate	Recall	Precision	Colour accuracy	Printed-board false positives
chip_test_v2 (UTH)	0.80	0.85	0.97	0 / 6 empty-board negatives
chip_test (blackjack)	0.92	0.95	—	—

The gen-1 printed-board false positive (spurious chip_black on the UTH "Play/Ante/Blind" ovals) is fixed here — gen 2 puts zero boxes on an empty printed board.

⚠️ Not casino accuracy (rule of the project). Gates are our own PoC recordings, never trained on. Colours only — mapping colour→denomination is per-casino configuration. Dataset-internal mAP@50 (0.70) in metrics.json is not a real-world accuracy claim.

Usage

import json, onnxruntime as ort
sess = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
colours = json.load(open("model.classes.json"))   # index -> "chip_black", ...
# Detect boxes per colour, then count per class for stack/bet estimation.

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.

Built for Live Game Defender (LGD) — an on-prem AI integrity monitor for live casino table games. © 2026 TechTools s.r.o.