Single-pass card boundary detector with 4-corner keypoint localization for Thai National ID cards. Designed for on-device inference at 30fps on mobile.

Metric	Value
Architecture	YOLO26n-Pose
Task	pose
Classes	1 (`thai-id-card`)
Keypoints	4 (TL, TR, BR, BL corners)
Box mAP50	0.993
Box mAP50-95	0.900
Pose mAP50	0.993
Pose mAP50-95	0.991
Parameters	~2.4M
GFLOPs	5.4 (fused)
Input	640x640
Training	Roboflow Card Detector v4 (303 images, 243 train / 60 val)
Experiment	EXP-001

Quick Start

python

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

model_path = hf_hub_download("chayuto/thai-id-ocr-yolo26n-card-corner-detector", "best.pt")
model = YOLO(model_path)
results = model("path/to/image.jpg")

Architecture

YOLO26n-Pose — NMS-free, DFL-free single-pass detector.

Input: [1, 3, 640, 640]
  → YOLO26n backbone + neck
  → 1-class bbox head (card boundary)
  → 4-keypoint pose head (TL, TR, BR, BL corners)
Output: [1, 300, 14]  (NMS-free, top-300 detections)
  → [x1, y1, x2, y2, conf, class, kp1_x, kp1_y, kp2_x, kp2_y, kp3_x, kp3_y, kp4_x, kp4_y]

Pipeline Context

This is the first stage of the Thai ID Nano OCR pipeline:

Camera Frame (640x640)
  → YOLO26n-Pose Finder (this model)
      → Card boundary + 4 corner keypoints
      → Perspective correction via corners
      → Zone extraction using fixed card layout (ISO 7810 ID-1)
  → 3 Reader Experts (Numeric, English, Thai)
  → Validator (CID Mod11 checksum, date logic)

The 4-corner keypoints enable perspective correction to normalize the card orientation before zone extraction. Since Thai ID cards follow the ISO 7810 ID-1 standard (85.60 x 53.98 mm), zone positions can be derived from the corner geometry without per-field detection.

TFLite Export

python

from ultralytics import YOLO
model = YOLO("best.pt")
model.export(format="tflite", int8=True, nms=False, imgsz=640)

Files

File	Description
`best.pt`	Ultralytics YOLO26n-pose weights (~5.6 MB)
`config.json`	Architecture params, training metadata, metrics

License

MIT