---
task_categories:
- visual-question-answering
- object-detection
language:
- en
tags:
- remote-sensing
- geospatial
- satellite-imagery
size_categories:
- 1K<n<10K
---

# GeoChat_bench_split Dataset

## Dataset Description

This dataset is a curated subset of the GeoChat_Bench benchmark from the EarthDial-Dataset, specifically containing samples with multiple ground truth bounding boxes.

### Dataset Summary

- **Total Samples:** 1,840
- **Source:** akshaydudhane/EarthDial-Dataset (GeoChat_Bench subset)
- **Task:** Visual grounding / object localization in satellite imagery
- **Modality:** Image (RGB) + Text

### Dataset Structure

The dataset contains 4 columns:

1. **question_id** (string): Unique identifier for each sample
2. **jpg** (Image): RGB satellite/aerial image
3. **question** (string): Question asking for object locations (format: "Give me the location of <ref>object description")
4. **groundtruth** (string): Multiple bounding box coordinates in the format `[[x1,y1,x2,y2,confidence]][[x1,y1,x2,y2,confidence]]...`

### Data Fields

- `question_id`: Unique sample identifier
- `jpg`: PIL Image object in RGB format
- `question`: Natural language question with reference tags
- `groundtruth`: Serialized list of bounding boxes

### Example

```python
{
    'question_id': 'sota_2403',
    'jpg': <PIL.Image>,
    'question': 'Give me the location of <ref>10 large small-vehicles',
    'groundtruth': '[[20, 43, 27, 53, 90]][[5, 38, 11, 47, 90]]...'
}
```

## Dataset Creation

### Curation Process

1. Loaded GeoChat_Bench from EarthDial-Dataset
2. Filtered samples containing multiple bounding boxes in ground truth (>1 list)
3. Sorted by number of bounding boxes in descending order
4. Selected top 1,840 samples (all available samples with multiple boxes)
5. Cleaned question text to remove formatting prefixes and tags

### Source Data

Original dataset: [EarthDial-Dataset](https://huggingface.co/datasets/akshaydudhane/EarthDial-Dataset)

## Usage

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("yobro4619/GeoChat_bench_split")

# Access samples
sample = dataset['train'][0]
print(sample['question'])
print(sample['groundtruth'])
```

## License

Please refer to the original [EarthDial-Dataset](https://huggingface.co/datasets/akshaydudhane/EarthDial-Dataset) for licensing information.

## Citation

If you use this dataset, please cite the original EarthDial-Dataset paper.