krutrim-ai-labs/ocr_rotation_bench

---
license: other
license_name: krutrim-community-license-agreement-version-1.0
license_link: LICENSE.md
task_categories:
- zero-shot-classification
language:
- en
- hi
- bn
- as
- ml
- mr
- ta
- pa
- or
- te
- kn
- gu
tags:
- image
- OCR
- rotation
- classification
pretty_name: seeingstraight
---

# ORB Rotation Benchmark

This repository contains the dataset for **“Seeing Straight: Document Orientation Detection for Efficient OCR”** (arXiv:2511.04161).

ORB Rotation Benchmark is a structured multi-domain dataset designed to evaluate document orientation detection models and their impact on downstream OCR performance.

Unlike prior document datasets that assume correct orientation, this benchmark explicitly introduces controlled rotation settings (4-class, 8-class, and 12-class configurations) across English and Indic document datasets, enabling systematic evaluation of rotation classification accuracy and OCR robustness.

The benchmark supports:

- Rotation classification benchmarking
- OCR performance evaluation with/without rotation correction
- Angle-wise error analysis
- Multilingual document evaluation

---

## Dataset Structure

The dataset consists of multiple splits corresponding to different language settings and rotation granularities.

### Splits

- `orb_en_4_class_sroie`
- `orb_en_4_class_synthdog`
- `orb_en_12_class`
- `orb_indic_4_class`
- `orb_indic_8_class`

Each split is stored as a Parquet file and is directly compatible with the Hugging Face Dataset Viewer.

---

## Data Format

Each dataset item contains:

| Field | Description |
|-------|------------|
| `image` | Document image |
| `rotation_label` | Rotation angle (integer degrees) |
| `ground_truth` | OCR ground-truth text (when available) |

### Rotation Settings

- 4-class: {0°, 90°, 180°, 270°}
- 8-class: evenly spaced rotations
- 12-class: finer-grained rotation classification

---

## Subset Description

### orb_en_4_class_sroie
English receipt dataset (SROIE-based) with 4 rotation classes and field-level OCR annotations.

### orb_en_4_class_synthdog
Synthetic English document dataset with 4 rotation classes and OCR ground truth.

### orb_en_12_class
English dataset with finer-grained 12-class rotation labels.

### orb_indic_4_class
Indic-language documents with 4 rotation classes.

### orb_indic_8_class
Indic-language documents with 8 rotation classes for higher granularity evaluation.

---

## Using the Dataset

Example usage:

```python
from datasets import load_dataset

ds = load_dataset("krutrim-ai-labs/orb_rotation_benchmark", split="orb_en_4_class_sroie")

print(ds[0])
```

## This dataset is intended for:

- Document orientation detection research

- OCR robustness evaluation

- Vision-language model benchmarking

- Multilingual document understanding

- Angle-wise performance analysis

- Evaluation Protocol

## The dataset enables:

- Rotation classification accuracy

- Angle-wise breakdown analysis

- OCR CER/WER before and after rotation correction

- Cross-language performance comparison

For full evaluation pipelines and experimental setup, refer to the paper.

## The official inference and evaluation pipeline is described in:

"Seeing Straight: Document Orientation Detection for Efficient OCR"

arXiv: https://arxiv.org/abs/2511.04161

### License

This repository is licensed under the Krutrim Community License.

### Citation

If you use this dataset, please cite:

```bibtex
@article{goswami2025seeing,
title={Seeing Straight: Document Orientation Detection for Efficient OCR},
author={Suranjan Goswami and Abhinav Ravi and Raja Kolla and Ali Faraz and Shaharukh Khan and Akash and Chandra Khatri and Shubham Agarwal},
year={2025},
eprint={2511.04161},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/pdf/2511.04161},
}
```