Blame for README.md

---
license: cc-by-nc-4.0
task_categories:
- visual-question-answering
- image-to-text
language:
- en
tags:
- blind
- low-vision
- accessibility
- egocentric
- navigation
- guidedog
size_categories:
- 10K<n<100K
gated: manual
extra_gated_heading: Request Access to GuideDog Dataset
extra_gated_prompt: 'This dataset is released for **non-commercial research purposes
  only**.


  By requesting access, you agree to:

  1. Use this dataset solely for academic research

  2. Not redistribute the dataset or any derived data

  3. Cite our work in any publications using this dataset

  4. Delete the dataset upon request from the authors


  Access requests are reviewed manually by the dataset authors.

  '
extra_gated_button_content: Submit Access Request
extra_gated_fields:
  Full Name: text
  Affiliation (University/Organization): text
  Country: country
  Email (institutional preferred): text
  Research Purpose:
    type: select
    options:
    - Academic Research
    - Education
    - Non-profit Research
    - Other
  Brief description of intended use: text
  I agree to use this dataset for non-commercial research only: checkbox
  I agree to cite the GuideDog paper in any publications: checkbox
configs:
- config_name: default
  data_files:
  - split: gold
    path: data/gold-*
  - split: silver
    path: data/silver-*
- config_name: depth
  data_files:
  - split: train
    path: depth/train-*
- config_name: object
  data_files:
  - split: train
    path: object/train-*
dataset_info:
- config_name: default
  features:
  - name: image
    dtype: image
  - name: gold_label
    dtype: string
  - name: silver_label
    dtype: string
  - name: scene_location
    dtype: string
  - name: scene_description
    dtype: string
  - name: video_country
    dtype: string
  - name: video_city
    dtype: string
  splits:
  - name: gold
    num_bytes: 254609775
    num_examples: 2106
  - name: silver
    num_bytes: 4446220487
    num_examples: 19978
  download_size: 4687100139
  dataset_size: 4700830262
- config_name: depth
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: choices
    list: string
  - name: closer_raw
    dtype: string
  - name: closer
    dtype: string
  - name: closer_bbox
    list: float64
  - name: closer_ratio_bbox
    list: float64
  - name: closer_direction
    dtype: string
  - name: closer_depth
    dtype: float64
  - name: closer_confidence
    dtype: float64
  - name: farther_raw
    dtype: string
  - name: farther
    dtype: string
  - name: farther_bbox
    list: float64
  - name: farther_ratio_bbox
    list: float64
  - name: farther_direction
    dtype: string
  - name: farther_depth
    dtype: float64
  - name: farther_confidence
    dtype: float64
  - name: video_country
    dtype: string
  - name: video_city
    dtype: string
  splits:
  - name: train
    num_bytes: 367866047
    num_examples: 383
  download_size: 367856018
  dataset_size: 367866047
- config_name: object
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: answer_raw
    dtype: string
  - name: answer
    dtype: string
  - name: answer_bbox
    list: float64
  - name: answer_ratio_bbox
    list: float64
  - name: answer_direction
    dtype: string
  - name: answer_depth
    dtype: float64
  - name: answer_confidence
    dtype: float64
  - name: choices
    list: string
  - name: video_country
    dtype: string
  - name: video_city
    dtype: string
  splits:
  - name: train
    num_bytes: 52715500
    num_examples: 435
  download_size: 50044669
  dataset_size: 52715500
---

# GuideDog Dataset

A benchmark dataset for blind and low-vision (BLV) navigation assistance with egocentric street-view images.

## Dataset Splits

| Split | Samples | Description |
|-------|---------|-------------|
| `gold` | ~2000 | Human-verified gold-standard annotations |
| `silver` | ~20000 | VLM-generated annotations for training |

**Usage:**
```python
from datasets import load_dataset

# Load gold split (for evaluation)
gold = load_dataset("kjunh/GuideDog", split="gold")

# Load silver split (for training)
silver = load_dataset("kjunh/GuideDog", split="silver")
```

## Features

| Column | Type | Description |
|--------|------|-------------|
| `image` | Image | Egocentric street-view image |
| `gold_label` | string | Human-verified guidance text (gold split only) |
| `silver_label` | string | VLM-generated guidance text (silver split only) |
| `scene_location` | string | Location type (e.g., sidewalk, crosswalk) |
| `scene_description` | string | Description of the scene |
| `video_country` | string | Country where video was recorded |
| `video_city` | string | City where video was recorded |

## Annotation Standards

Annotations follow the GuideDog S1/S2/S3 framework:
- **S1**: Scene description and user position context
- **S2**: Hazards at clock directions (10-2 o'clock) with distances
- **S3**: Summary and recommended navigation direction

## Intended Use

- **Gold split**: Evaluation and benchmarking
- **Silver split**: Training vision-language models

## Ethical Considerations

- All images are sourced from publicly available walking tour videos
- No personally identifiable information is included
- Dataset is provided for research purposes only

## Citation

If you use this dataset, please cite:

```bibtex
@misc{kim2025guidedogrealworldegocentricmultimodal,
      title={GuideDog: A Real-World Egocentric Multimodal Dataset for Blind and Low-Vision Accessibility-Aware Guidance},
      author={Junhyeok Kim and Jaewoo Park and Junhee Park and Sangeyl Lee and Jiwan Chung and Jisung Kim and Ji Hoon Joung and Youngjae Yu},
      year={2025},
      eprint={2503.12844},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.12844},
}
```

## License

This dataset is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).

## Contact

For questions or issues, please contact the dataset maintainers through the HuggingFace discussion tab.