Mean Dice

Available

instance_segmentation.mean_dice

Mean Dice overlap between matched predicted and ground-truth object instances.

Contract

Field	Value
version	1.0.0
metric_type	pointwise
scorer_contract	full_dataset
direction	higher_is_better
entrypoint	aip_scorers_instance_segmentation.answered.mean_dice
target_kind	None

Required columns

• image_id
• label
• predictions

Accepted schemas

[
  {
    "name": "gdi_image_v1",
    "task_types": [
      "instance_segmentation"
    ]
  }
]

Methodology

1. Decode per-object masks with one stable dataset class map.
2. Pair predictions to ground truth within each class, highest confidence first, keeping pairs at or above the match IoU threshold.
3. Score each matched pair by Dice; count every unmatched instance on either side as zero.
4. Average within each class, then across the classes the dataset observed.

Score semantics

Scores range from 0 to 1. Higher is better; 1 means every object was found once with a pixel-perfect mask.

Worked example

Two ground-truth objects matched at Dice 0.9 and 0.7, plus one missed object, give (0.9 + 0.7 + 0) / 3 = 0.53 for that class.

Configuration schema

{
  "properties": {
    "inverted": {
      "default": false,
      "title": "Inverted",
      "type": "boolean",
      "x-aip-param-role": "scoring_metadata"
    },
    "match_iou_threshold": {
      "default": 0.5,
      "description": "Minimum IoU at which a predicted instance is accepted as the same object as a ground-truth instance. Pairs below it leave both instances unmatched.",
      "maximum": 1.0,
      "minimum": 0.0,
      "title": "Match Iou Threshold",
      "type": "number"
    },
    "max": {
      "default": 1.0,
      "title": "Max",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "metric_name": {
      "default": "instance_segmentation_metric",
      "title": "Metric Name",
      "type": "string",
      "x-aip-param-role": "scoring_metadata"
    },
    "min": {
      "default": 0.0,
      "title": "Min",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "weight": {
      "default": 1.0,
      "title": "Weight",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    }
  },
  "title": "InitializationConfig",
  "type": "object"
}

Execution

{
  "emits_metric_family": false,
  "function_name": null,
  "max_concurrency": null,
  "processing_kind": "cpu",
  "stream_batch_size": null,
  "timeout_seconds": 600
}

Complete manifest

accepts:
- name: gdi_image_v1
  task_types:
  - instance_segmentation
config_schema:
  properties:
    inverted:
      default: false
      title: Inverted
      type: boolean
      x-aip-param-role: scoring_metadata
    match_iou_threshold:
      default: 0.5
      description: Minimum IoU at which a predicted instance is accepted as the same
        object as a ground-truth instance. Pairs below it leave both instances unmatched.
      maximum: 1.0
      minimum: 0.0
      title: Match Iou Threshold
      type: number
    max:
      default: 1.0
      title: Max
      type: number
      x-aip-param-role: scoring_metadata
    metric_name:
      default: instance_segmentation_metric
      title: Metric Name
      type: string
      x-aip-param-role: scoring_metadata
    min:
      default: 0.0
      title: Min
      type: number
      x-aip-param-role: scoring_metadata
    weight:
      default: 1.0
      title: Weight
      type: number
      x-aip-param-role: scoring_metadata
  title: InitializationConfig
  type: object
dependencies: []
description: Mean Dice overlap of matched object instances, averaged across classes.
direction: higher_is_better
display_name: Mean Dice
entrypoint: aip_scorers_instance_segmentation.answered.mean_dice
execution:
  emits_metric_family: false
  function_name: null
  max_concurrency: null
  processing_kind: cpu
  stream_batch_size: null
  timeout_seconds: 600
kind: metric
manifest_version: '1'
max_prompt_slots: 0
metric_metadata:
  methodology:
  - Decode per-object masks with one stable dataset class map.
  - Pair predictions to ground truth within each class, highest confidence first,
    keeping pairs at or above the match IoU threshold.
  - Score each matched pair by Dice; count every unmatched instance on either side
    as zero.
  - Average within each class, then across the classes the dataset observed.
  score_semantics: Scores range from 0 to 1. Higher is better; 1 means every object
    was found once with a pixel-perfect mask.
  summary: Mean Dice overlap between matched predicted and ground-truth object instances.
  worked_example: Two ground-truth objects matched at Dice 0.9 and 0.7, plus one missed
    object, give (0.9 + 0.7 + 0) / 3 = 0.53 for that class.
metric_type: pointwise
name: instance_segmentation.mean_dice
partition_types: []
required_columns:
- image_id
- label
- predictions
required_kinds: []
scorer_contract: full_dataset
target_kind: null
unsupported_trace_shapes: []
version: 1.0.0