yolo26n-video-detecter

YOLO26n Video Text Detector (Production v5)

Fine-tuned YOLO detector optimized for high-precision detection and spatial localization of burned-in video captions, dialogue subtitles, corner & floating watermarks, channel logos, and video text overlays.

---

1. Key Performance Highlights

Evaluated on the independent, strictly non-leaked 464-image validation benchmark (1,217 ground-truth instances):

1.1 Overall Metrics (Head-to-Head Comparison)

1.2 Class-by-Class Breakdown (mAP@50)

---

2. Canonical 4-Class Taxonomy

---

3. Quick Start & Inference

Installation

pip install ultralytics opencv-python

Python Inference Example

from ultralytics import YOLO

# Load the production checkpoint
model = YOLO('weights/best.pt')

# Run inference on image or video
results = model.predict(
    source='input_video.mp4',
    imgsz=1024,
    conf=0.25,
    device=0,
    save=True
)

for r in results:
    for box in r.boxes:
        cls_id = int(box.cls[0])
        conf = float(box.conf[0])
        coords = box.xyxy[0].tolist()
        print(f"Detected {model.names[cls_id]} ({conf:.2f}) at {coords}")

---

4. Training Configuration Details

• Dataset: 2,626 images (2,162 train, 464 validation) across 118 source videos (BnhHo/yolo26n-dataset).
• Image Resolution: 1024 x 1024 (preserves small fonts, diacritics, and semi-transparent watermarks).
• Optimizer: AdamW (lr0: 1e-4, weight_decay: 0.0005).
• Text-Safe Augmentations: fliplr=0.0, flipud=0.0, mosaic=0.0, degrees=0.0.
• Early Stopping: Epoch 18/60 (patience 7).

---

5. Citation

@model{yolo26n_video_detecter_2026,
  title={YOLO26n Burned-in Video Text Detector},
  author={BnhHo and MediaFlow Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/BnhHo/yolo26n-video-detecter}
}