---
license: agpl-3.0
datasets:
- Teklia/DAI-CReTDHI-RecordGold-ATR
base_model:
- Ultralytics/YOLO26
pipeline_tag: object-detection
---

# YOLOv26-m Record Detection

## Introduction

This version of YOLO26-m-OBB is specialized for record detection from double pages.
It was fine-tuned as part of the [DAI-CReTDHI](https://dai-cretdhi.univ-lr.fr/) project.

## Training

Parameters:
- Image extraction size: 2000 pixels
- Image training size: 1024 pixels
- Epochs: 189 epochs (early stopping)

## Evaluation


| Split | Images | Instances | Box(P | R     | mAP50 | mAP50-95) |
|:-----:|:------:|:---------:|:-----:|-------|-------|-----------|
| train | 1132   | 6549      | 0.985 | 0.985 | 0.994 | 0.95      |
| val   | 136    | 843       | 0.967 | 0.944 | 0.984 | 0.928     |
| test  | 139    | 807       | 0.965 | 0.979 | 0.991 | 0.92      |

### Usage

Use this model using `ultralytics`:

```sh
yolo detect predict model=model.pt source=path/to/images/
```