Shape-level Classification Example

Multi-task Classification

Introduction

Multi-task Classification involves training a model to perform multiple classification tasks simultaneously. For example, a model could be trained to classify both the type of person and vehicle attributes in a single image.

Usage

Step 0: Preparation

Prepare an attributes file such as attributes.json:

{
    "vehicle": {
        "bodyColor": [
            "red",
            "white",
            "blue"
        ],
        "vehicleType": [
            "SUV",
            "sedan",
            "bus",
            "truck"
        ]
    },
    "person": {
        "wearsGlasses": ["yes", "no"],
        "wearsMask": ["yes", "no"],
        "clothingColor": [
            "red",
            "green",
            "blue"
        ]
    },
    "__widget_types__": {
        "vehicle": {
            "bodyColor": "radiobutton",
            "vehicleType": "combobox"
        },
        "person": {
            "wearsGlasses": "radiobutton",
            "wearsMask": "radiobutton",
            "clothingColor": "combobox"
        }
    }
}

Note

Widget type configuration

You can specify the widget type for each attribute using the __widget_types__ section:

• "radiobutton": Single-click selection, ideal for yes/no or few options
• "group_id": Dropdown selection containing the group_ids of all objects, if any
• "lineedit": Free-form text input
• "combobox": Dropdown selection, suitable for many options (default)

If __widget_types__ is not specified, all attributes default to combobox.

Example of group_id and lineedit attributes:

{
    "vehicle": {
        "occluded_by": [],
        "occluded_by_string": ""
    },
    "__widget_types__": {
      "vehicle": {
        "occluded_by": "group_id",
        "occluded_by_string": "lineedit"
      }
    }
}

Step 1: Run the Application

python anylabeling/app.py

Step 2: Upload the Configuration File

Click on Upload -> Upload Attributes File in the top menu bar and select the prepared configuration file to upload.

Tip

Use Loop Select Shapes (Ctrl+Shift+C) to select shapes sequentially while assigning attributes.

Note

Attribute label color customization

You can customize the colors of attribute labels displayed on the canvas by configuring the .xanylabelingrc file in your user directory. Add or modify the following settings under the canvas section:

canvas:
  attributes:
    background_color: [33, 33, 33, 255]  # Background color (RGBA)
    border_color: [66, 66, 66, 255]      # Border color (RGBA)
    text_color: [33, 150, 243, 255]      # Text color (RGBA)

Color values use RGBA format: [R, G, B, A], where each value ranges from 0-255.

For detailed output examples, refer to this file.

Multiclass & Multilabel Classification

As with image-level classification, shape-level annotations support multiclass and multilabel flags.

Usage

GUI Import (Recommended)

Step 0: Preparation

Prepare a flags file like label_flags.yaml. An example is shown below:

person:
  - male
  - female
helmet:
  - white
  - red
  - blue
  - yellow
  - green

Step 1: Run the Application

python anylabeling/app.py

Step 2: Upload the Configuration File

Click on Upload -> Upload Label Flags File in the top menu bar and select the prepared configuration file to upload.

Command Line Loading

Option 1: Quick Start

python anylabeling/app.py --labels person,helmet --labelflags "{'person': ['male', 'female'], 'helmet': ['white', 'red', 'blue', 'yellow', 'green']}" --validatelabel exact

Tip

The labelflags key field supports regular expressions. For instance, you can use patterns like {person-\d+: [male, tall], "dog-\d+": [black, brown, white], .*: [occluded]}.

Option 2: Using a Configuration File

python anylabeling/app.py --labels labels.txt --labelflags label_flags.yaml --validatelabel exact

For detailed output examples, refer to this file.