---
license: apache-2.0
task_categories:
- object-detection
tags:
- scene-text
- text-detection
- video
- multilingual
---

# SceneText50

SceneText50 is a diverse multilingual outdoor dataset introduced in the paper [Back-Tracking from Clarity: Self-Learning to See Text from Afar](https://huggingface.co/papers/2609.32477).

It consists of 50 outdoor ego-centric videos (~840 clips) from 17 countries, with quadrilateral text annotations. The annotations are for **evaluation**; the self-learning pipeline introduced in the paper needs no manual labels for training.

Code: https://github.com/trid2912/BackTrackingText

## Dataset structure

```text
tracking_data/
├── 01/ 02/ ... 50/                 # one folder per video
│   ├── clip_10/                    # a clip = a short frame sequence
│   │   ├── clip_10_0000.jpg
│   │   ├── clip_10_0001.jpg
│   │   └── ...
│   ├── clip_10_0000.txt            # per-frame annotation (same basename as the frame)
│   ├── clip_10_0001.txt
│   └── ...
```

Each annotation line is one text instance:

```text
instance_id x1 y1 x2 y2 x3 y3 x4 y4
```

`instance_id` is the track ID (consistent across frames of a clip) and the four `(x, y)` pairs are the clockwise quadrilateral corners.

## Sample usage

Download from Hugging Face:

```bash
# option 1 — huggingface-cli
pip install -U "huggingface_hub[cli]"
huggingface-cli download Tri1/SceneText50 --repo-type dataset --local-dir data/tracking_data

# option 2 — git + git-lfs
git lfs install
git clone https://huggingface.co/datasets/Tri1/SceneText50 data/tracking_data
```

Point `--image_root` at this `tracking_data` folder when running the tracker / pipeline.