---
license: apache-2.0
---

# Instructions for Extracting Images from AndroidControl TFRecord

Follow these steps to download and extract images from the AndroidControl TFRecord dataset.

## Step 1: Download the Original AndroidControl TFRecord

1. Navigate to [the AndroidControl TFRecord repository](https://github.com/google-research/google-research/tree/master/android_control) where the AndroidControl TFRecord is hosted.
2. Download the TFRecord file to your local machine or server.

## Step 2: Run the Python Script to Extract Images

1. Ensure that `extract_images.py` is in the same directory as your downloaded TFRecord or update the script to point to the location of your TFRecord.
2. Open a terminal or command prompt.
3. Execute the script by running:
   ```bash
   python extract_images.py

## Step 3: Enjoy the parsed json data

### Here’s the information about the VH annotation:

Each screenshot corresponds to a `.pkl` file with the same name, containing a list of dictionaries. Each dictionary includes the following keys:

- **`"bounds_in_screen"`**: Defines the screen boundaries of the window, with values:
  ```json
  {
      "left": windows.bounds_in_screen.left,
      "right": windows.bounds_in_screen.right,
      "bottom": windows.bounds_in_screen.bottom,
      "top": windows.bounds_in_screen.top
  }
  ```
- **`"window_id"`**: The unique identifier of the window.
- **`"window_layer"`**: The layer or z-index of the window.
- **`"window_type"`**: The type of the window.
- **`"tree"`**: A list containing dictionaries representing each element node in the window.

Each node in the `"tree"` list includes:

- **`"bounds_in_screen"`**: The screen boundaries of the element node, with values:
  ```json
  {
      "left": node.bounds_in_screen.left,
      "right": node.bounds_in_screen.right,
      "bottom": node.bounds_in_screen.bottom,
      "top": node.bounds_in_screen.top
  }
  ```
- **`"class_name"`**: The class name of the element.
- **`"content_description"`**: The content description of the element.
- **`"package_name"`**: The package name the element belongs to.
- **`"text_selection_start"`**: The start position of any text selection.
- **`"text_selection_end"`**: The end position of any text selection.
- **`"view_id_resource_name"`**: The resource name of the view ID.
- **`"window_id"`**: The window ID associated with the element.
- **`"is_enabled"`**: Boolean indicating if the element is enabled.
- **`"is_visible_to_user"`**: Boolean indicating if the element is visible to the user.
- **`"actions"`**: A list of action IDs available for the element, derived as `[action.id for action in node.actions]`.
- **`"child_ids"`**: A list of IDs for child elements, derived as `[child_id for child_id in node.child_ids]`.
- **`"depth"`**: The depth of the element within the hierarchy.
- **`"drawing_order"`**: The order in which the element is drawn within the window.