Pixel Accuracy

Available

semantic_segmentation.pixel_accuracy

Per-class share of ground-truth pixels the predicted mask labels correctly.

Contract

Field	Value
version	1.0.0
metric_type	pointwise
scorer_contract	full_dataset
direction	higher_is_better
entrypoint	aip_scorers_semantic_segmentation.answered.pixel_accuracy
target_kind	None

Required columns

• image_id
• label
• predictions

Accepted schemas

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

Methodology

1. Decode canonical masks with one stable dataset class map.
2. Pool true-positive and ground-truth pixel counts per class across the whole dataset, then divide once.
3. Average the per-class values. Class id 0 covers pixels no annotation claims: it is excluded from the per-class artifact and from the macro average alike. A class the dataset never labels is left out of the average too, but is named in the artifact's unscored_classes list rather than dropped silently.

Score semantics

Scores range from 0 to 1. Higher is better; 1 means every ground-truth pixel of every class carries the right class.

Worked example

A class holding 5 ground-truth pixels of which 2 are predicted correctly scores 0.4, whichever images those pixels came from.

Configuration schema

{
  "description": "Configuration for the pooled per-class pixel-accuracy scorer.\n\nCarries no smoothing term: the denominator is a counted ground-truth area, and a\nclass whose area is zero is reported as unscored rather than smoothed towards a value.",
  "properties": {
    "inverted": {
      "default": false,
      "title": "Inverted",
      "type": "boolean",
      "x-aip-param-role": "scoring_metadata"
    },
    "max": {
      "default": 1.0,
      "title": "Max",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "metric_name": {
      "default": "pixel_accuracy",
      "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"
    },
    "threshold": {
      "default": 0.5,
      "maximum": 1.0,
      "minimum": 0.0,
      "title": "Threshold",
      "type": "number"
    },
    "weight": {
      "default": 1.0,
      "title": "Weight",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    }
  },
  "title": "PixelAccuracyInitializationConfig",
  "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:
  - semantic_segmentation
config_schema:
  description: 'Configuration for the pooled per-class pixel-accuracy scorer.


    Carries no smoothing term: the denominator is a counted ground-truth area, and
    a

    class whose area is zero is reported as unscored rather than smoothed towards
    a value.'
  properties:
    inverted:
      default: false
      title: Inverted
      type: boolean
      x-aip-param-role: scoring_metadata
    max:
      default: 1.0
      title: Max
      type: number
      x-aip-param-role: scoring_metadata
    metric_name:
      default: pixel_accuracy
      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
    threshold:
      default: 0.5
      maximum: 1.0
      minimum: 0.0
      title: Threshold
      type: number
    weight:
      default: 1.0
      title: Weight
      type: number
      x-aip-param-role: scoring_metadata
  title: PixelAccuracyInitializationConfig
  type: object
dependencies: []
description: Share of each class's ground-truth pixels the model labels correctly.
direction: higher_is_better
display_name: Pixel Accuracy
entrypoint: aip_scorers_semantic_segmentation.answered.pixel_accuracy
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'
metric_metadata:
  methodology:
  - Decode canonical masks with one stable dataset class map.
  - Pool true-positive and ground-truth pixel counts per class across the whole dataset,
    then divide once.
  - 'Average the per-class values. Class id 0 covers pixels no annotation claims:
    it is excluded from the per-class artifact and from the macro average alike. A
    class the dataset never labels is left out of the average too, but is named in
    the artifact''s unscored_classes list rather than dropped silently.'
  score_semantics: Scores range from 0 to 1. Higher is better; 1 means every ground-truth
    pixel of every class carries the right class.
  summary: Per-class share of ground-truth pixels the predicted mask labels correctly.
  worked_example: A class holding 5 ground-truth pixels of which 2 are predicted correctly
    scores 0.4, whichever images those pixels came from.
metric_type: pointwise
name: semantic_segmentation.pixel_accuracy
partition_types: []
required_columns:
- image_id
- label
- predictions
required_kinds: []
scorer_contract: full_dataset
target_kind: null
version: 1.0.0