Card Calibration v1

Inference weights for the Card Calibration project — automated color calibration via two-stage YOLO detection + tree-based regression.

Best result: XGBoost with Bayesian-tuned hyperparameters — Lab Mean ΔE = 4.59 (meets commercial printing standards).

Live Demo

Try it directly in your browser: HuggingFace Space

Files

File	Description	Size
`yolo_first.pt`	YOLOv8-nano — calibration card detector	6 MB
`yolo_second.pt`	YOLOv8-nano — 4-pattern detector (red/green/blue/black box)	6 MB
`xgboost_v1.pkl`	XGBoost calibration model (best)	3.5 MB
`random_forest_v1.pkl`	Random Forest calibration model	45 MB

Quick Start

from huggingface_hub import hf_hub_download
import joblib
from ultralytics import YOLO

REPO = "jeffliulab/card-calibration-v1"

# Download weights (cached in ~/.cache/huggingface/)
yolo_card_path    = hf_hub_download(repo_id=REPO, filename="yolo_first.pt")
yolo_pattern_path = hf_hub_download(repo_id=REPO, filename="yolo_second.pt")
xgb_path          = hf_hub_download(repo_id=REPO, filename="xgboost_v1.pkl")

# Load
yolo_card    = YOLO(yolo_card_path)
yolo_pattern = YOLO(yolo_pattern_path)
xgb_model    = joblib.load(xgb_path)

For the full inference pipeline (detect card → detect patterns → extract RGB → predict true color), see space/model_utils.py.

Inference Pipeline

Photo ──▶ YOLO Stage 1 ──▶ Card Crop ──▶ YOLO Stage 2 ──▶ 4 Patches
                                                              │
                                          ┌───────────────────┘
                                          ▼
                                   Feature Engineering
                                   (9 deltas + 3 RGB)
                                          │
                                          ▼
                                   XGBoost / RF Model
                                          │
                                          ▼
                                   Predicted True RGB

Performance

Model	R²	RMSE	Lab Mean ΔE	Lab Median ΔE
XGBoost (tuned)	0.8280	11.76	4.59	3.61
Random Forest	0.8225	12.10	5.20	3.96

ΔE < 3: professional · ΔE < 5: commercial printing · ΔE < 10: acceptable

Training Data

Dataset: jeffliulab/card-calibration-v1-data

255 hand-collected photos of color calibration cards
Augmented to 2,294 samples (brightness/hue/blur/noise/rotation)
12-D feature vectors extracted from center 1/3 of each detected patch
70/30 train/test split, random_state=42

License

MIT — both data and model weights are freely available for research and commercial use.

Downloads last month: 36

jeffliulab
/

card-calibration-v1