---
license: cc-by-nc-sa-4.0
---
---
license: cc-by-nc-sa-4.0
---

# EMT Dataset

This dataset was presented in [EMT: A Visual Multi-Task Benchmark Dataset for Autonomous Driving in the Arab Gulf Region](https://huggingface.co/papers/2502.19260).

## Introduction
EMT is a comprehensive dataset for autonomous driving research, containing **57 minutes** of diverse urban traffic footage from the **Gulf Region**. It includes rich semantic annotations across two agent categories:

- **People**: Pedestrians and cyclists
- **Vehicles**: Seven different classes

Each video segment spans **2.5-3 minutes**, capturing challenging real-world scenarios:

- **Dense Urban Traffic** – Multi-agent interactions in congested environments
- **Weather Variations** – Clear and rainy conditions
- **Visual Challenges** – High reflections and adverse weather combinations (e.g., rainy nights)

### Dataset Annotations
This dataset provides annotations for:

- **Detection & Tracking** – Multi-object tracking with consistent IDs

For **intention prediction** and **trajectory prediction** annotations, please refer to our [GitHub repository](https://github.com/AV-Lab/emt-dataset).

---

## Quick Start
```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("KuAvLab/EMT", split="train")
```

### Available Labels
Each dataset sample contains two main components:

1. **Image** – The frame image
2. **Object** – The annotations for detected objects

#### Object Labels
- **bbox**: Bounding box coordinates (`x_min, y_min, x_max, y_max`)
- **track_id**: Tracking ID of detected objects
- **class_id**: Numeric class ID
- **class_name**: Object type (e.g., `car`, `pedestrian`)

#### Sample Usage
```python
import numpy as np

for data in dataset:
    # Convert image from PIL to OpenCV format (BGR)
    img = np.array(data['image'])
    
    print("Classes:", data['objects']['class_name'])
    print("Bboxes:",  data['objects']['bbox'])
    print("Track IDs:", data['objects']['track_id'])
    print("Class IDs:", data['objects']['class_id'])
```

---

## Data Collection
| Aspect      | Description                        |
|------------|----------------------------------|
| Duration   | 57 minutes total footage        |
| Segments   | 2.5-3 minutes per recording     |
| FPS        | 10 fps for annotated frames     |
| Agent Classes | 2 Person categories, 7 Vehicle categories |

### Agent Categories
#### **People**
- Pedestrians
- Cyclists

#### **Vehicles**
- Motorbike
- Small motorized vehicle
- Medium vehicle
- Large vehicle
- Car
- Bus
- Emergency vehicle

---

## Dataset Statistics
| Category           | Count       |
|-------------------|------------|
| Annotated Frames  | 34,386     |
| Bounding Boxes    | 626,634    |
| Unique Agents     | 9,094      |
| Vehicle Instances | 7,857      |
| Pedestrian Instances | 568  |

### Class Breakdown
| **Class**                 | **Description** | **Bounding Boxes** | **Unique Agents** |
|---------------------------|----------------|-------------------|----------------|
| Pedestrian                | Walking individuals  | 24,574 | 568 |
| Cyclist                   | Bicycle/e-bike riders  | 594 | 14 |
| Motorbike                 | Motorcycles, bikes, scooters  | 11,294 | 159 |
| Car                       | Standard automobiles | 429,705 | 6,559 |
| Small motorized vehicle   | Mobility scooters, quad bikes  | 767 | 13 |
| Medium vehicle            | Vans, tractors  | 51,257 | 741 |
| Large vehicle             | Lorries, trucks (6+ wheels) | 37,757 | 579 |
| Bus                       | School buses, single/double-deckers  | 19,244 | 200 |
| Emergency vehicle         | Ambulances, police cars, fire trucks  | 1,182 | 9 |
| **Overall**               | | **576,374** | **8,842** |

---

For more details , visit our [GitHub repository](https://github.com/AV-Lab/emt-dataset).