---
language:
- en
task_categories:
- object-detection
tags:
- computer-vision
- object-detection
- bicycle
- grayscale
- embedded-ai
- tinyml
pretty_name: Bike Detection Subset
size_categories:
- n<1K
---

# Bike Detection Subset

## Dataset Description

This dataset is a small public subset of a private bicycle detection dataset used for embedded computer vision experiments.

It contains **200 grayscale QQVGA images**:

- 100 positive samples containing at least one bicycle
- 100 negative samples without bicycles

The images have a resolution of:

```text
160 × 120
```

and are stored as grayscale images.

## Dataset Structure

Each sample contains:

```text
image
image_id
task
sample_type
width
height
objects
```

`sample_type` is either:

```text
positive
negative
```

The `objects` field contains:

```text
bbox
category
```

Bounding boxes use the format:

```text
[x, y, width, height]
```

with coordinates expressed in pixels.

The bicycle class is represented with category ID:

```text
0
```

Negative samples contain no bounding boxes.

## Example

```python
{
    "image": ...,
    "image_id": "task_01_image_001",
    "task": "task_01",
    "sample_type": "positive",
    "width": 160,
    "height": 120,
    "objects": {
        "bbox": [
            [34.0, 21.0, 48.0, 71.0]
        ],
        "category": [
            0
        ]
    }
}
```

## Dataset Creation

The samples were randomly selected from a larger private dataset.

The original data consists of low-resolution grayscale images prepared for an embedded bicycle detection use case.

Annotations were originally stored in YOLO format and converted to pixel-space bounding boxes when generating this Hugging Face dataset.

## Intended Use

This subset is intended for:

- Demonstrating the associated bicycle detection model
- Testing dataset loading and preprocessing pipelines
- Embedded computer vision experiments
- Bicycle object detection examples

It is not intended to be a comprehensive benchmark dataset.

## Limitations

This dataset contains only 200 samples and represents a limited subset of the original private dataset.

It may not cover the full diversity of:

- Lighting conditions
- Camera viewpoints
- Bicycle appearances
- Backgrounds
- Object scales
- Acquisition environments

Performance measured on this subset should therefore not be considered representative of general bicycle detection performance.

## Source Dataset

The complete training dataset is private and is not distributed with this repository.