walidhadri/table-detection-yolo26n

Table Detection YOLOv26n (TableBank)

Model Details

This model is a YOLOv26n-based detector trained to locate tables in document images.
It was trained with a simple setup (no hyperparameter tuning or model-size changes) and already gives strong results. Further gains are likely possible with more tuning.

Training Data

• Dataset: TableBank (document table images)

Training Procedure

• Architecture: YOLOv26n
• Epochs: ~10 (training logs include 9 epochs)
• Notes: Minimal tuning; default-style training was sufficient for good results.

Results (from results.csv)

Final logged epoch (epoch 9) validation metrics:

• Precision: 0.9519
• Recall: 0.9604
• mAP@0.5: 0.9865
• mAP@0.5:0.95: 0.9716

Intended Use

• Detecting tables in scanned or digital document images.

Limitations

• Trained only on TableBank; performance may drop on very different layouts or document styles.
• No hyperparameter tuning performed yet; results can likely be improved.

How to Use

Load with Ultralytics YOLO:

from ultralytics import YOLO
model = YOLO("yolo26n-tablebank.pt")
results = model("path/to/image.jpg")
results[0].show()


---
license: mit
---