ThreeSixNine/poker-table-perception-eval

Poker-Perception Eval · synthetic tables

Synthetic 1600×900 poker table screenshots rendered deterministically (seed 17), each paired with exact ground-truth labels, built to benchmark how well small on-device vision models can read a poker table without touching any real-money client. All pixels are programmatically generated — no game footage, no ToS surface, fully redistributable.

Fine-tune YOLO on these tablesBuild the decision engine

Contents

• data/images/test_000.png … test_059.png — 1600×900 renders: green felt, 2–6 seats (random), face-up hole + hero cards, flop/turn/river board subset, pot text, per-seat stack/position plates, occasional bet circles, dealer button.
• data/labels/test_NNN.json — exact ground truth: cards ("A♠" style), pixel bboxes per card, per-seat stacks/positions/optional bets, pot, dealer seat. 741 labeled card instances total (avg 12.3/image).
• render_tables.py — the generator (deterministic; regenerate with --seed).
• eval_mirage.py, eval_yolo.py — the two scorers used in the eval below.
• eval/results_mirage_1x.json, eval/results_mirage_2x.json, results_yolo.json — raw per-image outputs of the phase-1 eval. eval/results_stock_test.json, eval/results_ft_test.json — before/after eval of the phase-2 fine-tune on this same seed-17 test set.
• benchmark/ — the phase-2 training benchmark: 800 train (seed 101) + 80 val (seed 202) varied-style images in YOLO format (images/, labels/, data.yaml) with full ground-truth JSONs under gt/. The seed-17 set is never in training.
• train_yolo.py — the fine-tune script the GPU job ran.

Eval results (CPU, 2 vCPU)

Model	Card read / cls acc	Localization	Pot text	s/img
Mirage 51M (Cactus, CPU)	0.0	best-IoU ≈ 0.04	0.97	6.4
Mirage @ 2× upscale	0.0	0.005	1.00	6.7
YOLO11n cards (shrimantasatpati, MIT, conf 0.1, imgsz 1280)	0.119 (glyph region)	0.193 (glyph region IoU≥0.4); 0.0 vs full-card box	n/a	0.37

Findings:

1. Mirage OCR is a table-state reader, not a card reader. It recovers pot, stacks, names, positions reliably, but produces no card-glyph text at either resolution — and 2× upscale doesn't fix it (localization gets worse).
2. The YOLO card detector detects corner glyph patches, not whole cards on our renders; scoring against full-card boxes yields 0.0 even though ~28% of cards have a centered detection. Against the corner-glyph region its detection rate is 19.3% with 11.9% correct class — weak, because its real-photo training set is off-distribution for flat synthetic renders.
3. Implication for the stack: read state (pot/stacks) with Mirage-class OCR; read card identity with a detector fine-tuned on this synthetic data (or your captured-game frames later). The two failures are complementary and separable.

Fine-tune results (phase 2)

YOLO11n fine-tuned 80 epochs (batch 16, imgsz 1280) on benchmark/ → ThreeSixNine/yolo11n-poker-table-cards. Scored on the untouched seed-17 test set (741 cards):

Model	Detection (IoU≥0.4)	Class accuracy	s/img
Stock	0.193	0.119	0.36
Fine-tuned	0.999	0.586	0.33

5× class accuracy, essentially perfect localization, no inference cost. Class accuracy below 1.0 is 52-way rank/suit confusion on unseen domain detail; per-image worst 6/17, best 12/12. Run cost ≈ $0.53 on a10g-small (31.6 min).

Sim-to-real results (phase 3)

Real-photo eval at imgsz 224 (scale-matched to these sets):

Real set	Stock det / cls	Fine-tuned det / cls
gpiosenka crops (520 imgs, CC0)	0.000 / 0.000	0.571 / 0.329
gunhcolab XML-box scenes (98 imgs)	0.000 / 0.000	0.929 / 0.316

• Stock never emits full-card boxes on real photos at any scale (corner-pip patches only) — 0.0 across the board.
• The fine-tuned model transfers detection well (0.93 on real full scenes at matched scale) but class accuracy drops 0.586 → ~0.32 — a ~26-point sim-to-real gap concentrated in the classify head, with failures skewed toward spades and queens/threes.
• Inference-scale lesson: at imgsz 1280 both models score 0.0 on these sets — an artifact of letterboxing 224×224 crops to the synthetic-table scale. imgsz is part of the model contract and must match the data scale.
• gunhcolab files are NOT redistributed (license © original authors); only aggregate metrics above are published. gpiosenka-derived numbers are CC0-safe.

v2 mixed fine-tune: head-to-head (phase 4)

benchmark_2/ — 800+80 scenes, ~40% of card slots pasted with real gpiosenka TRAIN-split crops at table scale (CC0; valid/test untouched). Fine-tuned from the phase-2 checkpoint → best_mixed.pt in ThreeSixNine/yolo11n-poker-table-cards:

Eval	FT-v1	FT-mixed
Synthetic seed-17 (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 doubles scene-level real class accuracy with no synthetic regression. Raw outputs: eval/{syn_ft1,syn_ftm,real_gp_*,real_gu_*}.json.

Reproduce

python render_tables.py --n 60 --seed 17
python eval_yolo.py --limit 60            # conf 0.1, imgsz 1280 defaults
python eval_mirage.py --limit 60          # cactus-compute==4.0.1