Instructions to use ThreeSixNine/yolo11n-poker-table-cards with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- ultralytics
How to use ThreeSixNine/yolo11n-poker-table-cards with ultralytics:
from huggingface_hub import hf_hub_download from ultralytics import YOLO # pick the weights file from this repo's "Files and versions" tab weights = hf_hub_download("ThreeSixNine/yolo11n-poker-table-cards", "<weights>.pt") model = YOLO(weights) source = 'http://images.cocodataset.org/val2017/000000039769.jpg' model.predict(source=source, save=True) - Notebooks
- Google Colab
- Kaggle
yolo11n-poker-table-cards
YOLO11n card detector fine-tuned to read face-up cards on synthetic poker-table screenshots (1600ร900, varied skins/scales/fonts/noise), built from shrimantasatpati/yolov11_playing_cards_detection (MIT base) and the synthetic benchmark in ThreeSixNine/poker-table-perception-eval: 800 train + 80 val renders, 52 classes, class order identical to the base checkpoint.
Results (untouched seed-17 test set, 741 cards, conf 0.1, imgsz 1280)
| Model | Detection IoUโฅ0.4 | Class accuracy | s/img (CPU) |
|---|---|---|---|
| Base checkpoint | 0.193 | 0.119 | 0.36 |
| This fine-tune | 0.999 | 0.586 | 0.33 |
Localization is essentially solved on synthetic tables; class accuracy is bounded by 52-way rank/suit confusion. Real-world (photographed/Steam-capture) transfer is untested โ that's the next experiment, and real-card photos would quantify it.
v2 fine-tune (mixed cut-paste)
best_mixed.pt continues from best.pt on benchmark_2 in the dataset repo:
800+80 table scenes where ~40% of card slots are REAL card crops (gpiosenka
train split, CC0) pasted at table scale with brightness/contrast/blur jitter.
Head-to-head on untouched eval sets:
| Eval | FT-v1 (best.pt) |
FT-mixed (best_mixed.pt) |
|---|---|---|
| Synthetic seed-17 test (det / cls) | 0.999 / 0.586 | 0.996 / 0.594 |
| gpiosenka crops (det / cls) | 0.571 / 0.329 | 0.650 / 0.469 |
| gunhcolab scenes (det / cls) | 0.929 / 0.316 | 0.990 / 0.735 |
Cut-paste sim2real roughly doubles real-photo class accuracy on scenes (0.316 โ 0.735) with no synthetic regression. Residual gap on isolated crops (0.469) reflects the scene-context mismatch. Run: 31 min a10g-small (~$0.52).
Usage
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
w = hf_hub_download("ThreeSixNine/yolo11n-poker-table-cards", "best.pt")
model = YOLO(w)
r = model.predict("table.png", conf=0.1, imgsz=1280)
Training
80 epochs, batch 16, imgsz 1280, a10g-small (~$0.53, 31.6 min). Script:
train_yolo.py in the dataset repo.
- Downloads last month
- 140