If you are training an object detector, you need bounding boxes. This tutorial starts with a minimal setup and builds up to attributes, validation rules, and a full production config. For the image options behind it, see the image annotation documentation.

Basic Bounding Box Setup

Minimal Configuration

yaml

annotation_task_name: "Object Detection Annotation"
 
data_files:
  - "data/images.json"
 
annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    description: "Draw boxes around all objects"
    tools:
      - bbox
    labels:
      - name: car
        color: "#FF6B6B"
      - name: person
        color: "#4ECDC4"
      - name: bicycle
        color: "#45B7D1"

How It Works

Select a label from the toolbar
Click and drag to draw a box
Adjust corners to refine the box
Add more boxes as needed
Submit when done

Annotators draw bounding boxes directly on images with a click-and-drag interface:

Bounding box annotation interface for object detection The bounding box tool lets annotators draw and label rectangular regions on images

Detailed Label Configuration

yaml

annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    description: "Annotate all visible objects"
    tools:
      - bbox
    labels:
      - name: person
        color: "#FF6B6B"
        description: "Any human, partial or full"
        keyboard_shortcut: "p"
 
      - name: car
        color: "#4ECDC4"
        description: "Cars, trucks, SUVs"
        keyboard_shortcut: "c"
 
      - name: motorcycle
        color: "#45B7D1"
        description: "Motorcycles and scooters"
        keyboard_shortcut: "m"
 
      - name: bicycle
        color: "#96CEB4"
        description: "Bicycles of all types"
        keyboard_shortcut: "b"
 
      - name: traffic_light
        color: "#FFEAA7"
        description: "Traffic signals"
        keyboard_shortcut: "t"
 
      - name: stop_sign
        color: "#DDA0DD"
        description: "Stop signs"
        keyboard_shortcut: "s"

Adding Object Attributes

Capture additional information about each box:

yaml

annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    tools:
      - bbox
    labels:
      - name: person
        color: "#FF6B6B"
        attributes:
          - name: occlusion
            type: radio
            options: [none, partial, heavy]
          - name: truncated
            type: checkbox
            description: "Object extends beyond image"
          - name: difficult
            type: checkbox
            description: "Hard to identify"

When annotators draw a box, they'll be prompted to fill in these attributes.

Validation Rules

Ensure annotation quality with validation:

yaml

annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    tools:
      - bbox
    labels: [...]
    min_annotations: 1

Keyboard Shortcuts

Potato includes built-in keyboard shortcuts for efficient annotation:

Number keys select labels
Delete removes selected annotations
Arrow keys navigate between items

Display Options

Configure image display settings:

yaml

image_display:
  width: 800
  height: 600

Complete Production Configuration

yaml

annotation_task_name: "Autonomous Driving - Object Detection"
 
data_files:
  - "data/driving_frames.json"
 
annotation_schemes:
  - annotation_type: image_annotation
    name: objects
    description: "Annotate all traffic participants and objects"
    tools:
      - bbox
    min_annotations: 1
    labels:
      - name: vehicle
        color: "#FF6B6B"
        keyboard_shortcut: "v"
        attributes:
          - name: type
            type: radio
            options: [car, truck, bus, motorcycle, bicycle]
          - name: occlusion
            type: radio
            options: [0%, 1-25%, 26-50%, 51-75%, 76-99%]
 
      - name: pedestrian
        color: "#4ECDC4"
        keyboard_shortcut: "p"
        attributes:
          - name: pose
            type: radio
            options: [standing, walking, sitting, lying]
          - name: age_group
            type: radio
            options: [child, adult, elderly]
 
      - name: cyclist
        color: "#45B7D1"
        keyboard_shortcut: "c"
 
      - name: traffic_sign
        color: "#FFEAA7"
        keyboard_shortcut: "t"
        attributes:
          - name: sign_type
            type: radio
            options: [stop, yield, speed_limit, warning, other]
 
      - name: traffic_light
        color: "#DDA0DD"
        keyboard_shortcut: "l"
        attributes:
          - name: state
            type: radio
            options: [red, yellow, green, off, unknown]

Output Format

json

{
  "frame_id": "frame_0001",
  "frame_path": "/images/frame_0001.jpg",
  "image_dimensions": {"width": 1920, "height": 1080},
  "annotations": {
    "objects": [
      {
        "label": "vehicle",
        "bbox": [450, 380, 680, 520],
        "attributes": {
          "type": "car",
          "occlusion": "0%"
        }
      },
      {
        "label": "pedestrian",
        "bbox": [820, 400, 870, 550],
        "attributes": {
          "pose": "walking",
          "age_group": "adult"
        }
      }
    ]
  }
}

Tips for Quality Bounding Boxes

Draw boxes tight to the object so you are not feeding the model a lot of background. Write down how to handle the awkward cases (occlusion, objects cut off at the edge) and review a batch of examples together early on so everyone draws them the same way. Keep an eye on box counts and size distributions; a sudden shift usually means someone is interpreting the rules differently.

Next Steps

Add polygon annotation for precise segmentation
Learn about keypoint detection for pose estimation
Set up quality control for reliable annotations

Full documentation at /docs/features/image-annotation.