> For the complete documentation index, see [llms.txt](https://docs.augelab.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.augelab.com/key-features/annotate-data-for-object-detection/dataset-collection.md).

# Dataset Collection

The fastest way to build a high-performance AI model is to **capture data on purpose**. This page covers how to collect high-quality images and videos using AugeLab Studio's native tools.

{% hint style="info" %}
You may skip this section if you already have a folder of images/videos ready for annotation.
{% endhint %}

***

## Planning Your Dataset

It's crucial to plan your dataset before collection. A well-structured dataset leads to better model performance.

### 📊 How Much Data Do You Need?

The number of images required depends on how much the environment changes. Use this table as a starting point for your collection goal.

| Project Type        | Environment                                     | Recommended Images per class\* |
| ------------------- | ----------------------------------------------- | ------------------------------ |
| **Simple**          | Controlled lighting, fixed camera, 1-2 classes. | **50 - 150**                   |
| **Industrial**      | Factory floor, changing shifts, conveyor belt.  | **200 - 500**                  |
| **Complex**         | Variable lighting, many classes, moving camera. | **1,000+**                     |
| **Complex Outdoor** | Outdoor scenes with weather changes.            | **2,000+**                     |
| **Rare Event**      | Detecting occasional defects or leaks.          | **50 Target / 100 Empty**      |

> \*Images per class refers to the number of annotated instances of each object category, not just total images.

{% hint style="info" %}
For best results, aim for **diversity** in angles, distances, and lighting within your dataset.
{% endhint %}

{% hint style="warning" %}
Number of classes should be consistent across the dataset. If not, **augmentation** can help balance classes later.
{% endhint %}

***

### 🏗️ Define the boundaries:

Write these down before taking the first photo to ensure your dataset is **Representative** and **Consistent**.

1. **Class List**: What specific objects are you detecting?
2. **Camera Specs**: What is the final mounting angle, distance, and Field of View (FoV)? Single or multiple cameras?
3. **Variations**: Will there be shifts in lighting (glare/shadows) or background clutter?
4. **Negatives**: What does an "empty" scene look like?
5. **Scope**: What objects should the model intentionally ignore?

***

## Camera Configuration

Whether using a USB camera, IP camera, or industrial camera, ensure the following settings are optimized before collection:

* **Resolution**: Aim for 480p to 720p (640x480 is a common standard). Higher resolutions can be downscaled later.
* **Frame Rate**: 15-30 FPS is sufficient for most object detection tasks.
* **Focus**: Set to manual focus to avoid shifts during collection.
* **Exposure**: Use manual exposure settings to maintain consistent lighting.
* **Save Settings**: Save your camera settings profile, most cameras allow saving presets, so the settings remain consistent across sessions.

## Dataset Collection

You can collect images and videos for your object detection dataset directly within AugeLab Studio using built-in tools. This ensures compatibility and streamlines the annotation process.

> Another option is to download public datasets or use external cameras/software, but this may require additional formatting steps.

### Capture Inside AugeLab Studio

Using the Studio environment allows you to use triggers (buttons, PLC signals, or timers) to automate your collection.

### 1. Start from the Example Project

AugeLab ships with a pre-configured template for this exact task.

* **Path**: `File` → `Example Projects` (or "Example Scenarios")
* **Project**: **"Data Collection for AI Training"**

### 📸 Single Images: The `Image Write` Block

Use this for high-quality static frames. It is best for "same scene, many positions."

| Input/Setting      | Logic                                                                 |
| ------------------ | --------------------------------------------------------------------- |
| **Folder Path**    | Where images are stored.                                              |
| **Save (Trigger)** | Set to `True` to capture a frame. Pair this with a button or a timer. |
| **Compress Image** | **Checked** = `.jpg` (Smaller)                                        |

### 🎥 Continuous Motion: The `Record Video` Block

Best for conveyor belts or fast-moving inspections where you intend to extract frames later.

| Input/Setting              | Logic                                   |
| -------------------------- | --------------------------------------- |
| **Video Quality**          | **Compressed** = `.mp4`                 |
| **Trigger Mode: Spacebar** | Press Space to Start/Stop.              |
| **Trigger Mode: Once**     | `Record=True` toggles recording on/off. |

> Plan recordings as short, focused clips (10–60s) rather than one massive file. This makes frame extraction much easier.

***

## 📉 Collecting Background (Negative) Images

A robust model needs to know what *not* to detect. You must capture "Empty" scenes on purpose.

* **What to capture**: Empty conveyors, empty workstations, or common non-target objects (fixtures, tools).
* **Empty**: An annotation file exists, but has no boxes.
* **Excluded**: No annotation file exists.

***

## Public Datasets

If you need to supplement your own data, consider these public datasets:

* [COCO Dataset](https://cocodataset.org/#home): Large-scale object detection, segmentation, and captioning dataset.
* [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/): Standard dataset for object detection and segmentation.
* [Open Images Dataset](https://storage.googleapis.com/openimages/web/index.html): A dataset with \~9 million images annotated with image-level labels and bounding boxes.
* [ImageNet](http://www.image-net.org/): Large visual database designed for use in visual object recognition research.
* [Kaggle Datasets](https://www.kaggle.com/datasets): Various datasets for machine learning, including object detection.

## 📂 Folder Structure & Preparation

AugeLab Studio loads datasets by folder. Ensure your structure looks like this:

```
my_dataset/
  ├── 000001.jpg
  ├── 000002.jpg
  ├── background_01.png
  └── classes.names  <-- (Optional, will be created during annotation)
```

***

## 🏁 Capture Checklist

| Check          | Requirement                                                       |
| -------------- | ----------------------------------------------------------------- |
| **Quality**    | Avoid heavy motion blur or over-exposure where edges disappear.   |
| **Coverage**   | Capture objects in the center, corners, and edges of the frame.   |
| **Scale**      | Match the real-world distance from the camera to the object.      |
| **Clutter**    | Include the messy backgrounds the camera will actually see.       |
| **Resolution** | Most AI models work best between 480p and 720p (640x480 average). |