MSAD Multi-Scenario Anomaly Detection

# MSAD Multi-Scenario Anomaly Detection Configuration
# Based on Zhang et al., NeurIPS 2024
# Task: Detect and classify anomalous events in surveillance-style videos

annotation_task_name: "MSAD Multi-Scenario Anomaly Detection"
task_dir: "."

# Data configuration
data_files:
  - sample-data.json
item_properties:
  id_key: "id"
  text_key: "video_url"

# Output
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

# Annotation schemes
annotation_schemes:
  # Temporal anomaly segment marking
  - name: "anomaly_segments"
    description: |
      Mark temporal segments where anomalous events occur.
      Drag to select the start and end of each anomalous event.
      Normal segments do not need to be marked.
    annotation_type: "video_annotation"
    mode: "segment"
    labels:
      - name: "anomaly"
        color: "#EF4444"
        key_value: "a"
      - name: "suspicious"
        color: "#F59E0B"
        key_value: "s"
    frame_stepping: true
    show_timecode: true
    playback_rate_control: true
    zoom_enabled: true
    timeline_height: 80

  # Anomaly type classification
  - name: "anomaly_type"
    description: |
      Classify the type of anomaly detected in the marked segment.
      Select the most appropriate category for the anomalous event.
    annotation_type: radio
    labels:
      - name: "Violence"
        tooltip: "Physical altercation, fighting, assault"
        key_value: "1"
      - name: "Theft"
        tooltip: "Stealing, pickpocketing, shoplifting, robbery"
        key_value: "2"
      - name: "Traffic Violation"
        tooltip: "Running red lights, wrong-way driving, reckless driving"
        key_value: "3"
      - name: "Fire"
        tooltip: "Fire, smoke, explosion"
        key_value: "4"
      - name: "Vandalism"
        tooltip: "Property damage, graffiti, destruction"
        key_value: "5"
      - name: "Trespassing"
        tooltip: "Unauthorized entry, breaking and entering, fence climbing"
        key_value: "6"
      - name: "Normal"
        tooltip: "No anomaly detected, normal activity"
        key_value: "7"

  # Severity assessment
  - name: "severity"
    description: "Rate the severity of the detected anomaly"
    annotation_type: radio
    labels:
      - name: "Low"
        tooltip: "Minor incident, low risk"
        key_value: "q"
      - name: "Medium"
        tooltip: "Moderate incident, potential risk"
        key_value: "w"
      - name: "High"
        tooltip: "Serious incident, significant risk"
        key_value: "e"
      - name: "Critical"
        tooltip: "Severe incident, immediate danger"
        key_value: "r"

# User configuration
allow_all_users: true

# Task assignment
instances_per_annotator: 40
annotation_per_instance: 3

# Instructions
annotation_instructions: |
  ## Multi-Scenario Video Anomaly Detection Task

  Your goal is to detect and classify anomalous events in surveillance-style video clips.

  ### Step 1: Watch the Video
  - Watch the entire video at normal speed first
  - Note any unusual or abnormal events

  ### Step 2: Mark Anomaly Segments
  - Use the timeline to mark the START and END of each anomalous event
  - Mark clearly anomalous events as "anomaly" (red)
  - Mark borderline/uncertain events as "suspicious" (yellow)
  - Normal segments do not need annotation

  ### Step 3: Classify Anomaly Type
  For each marked segment, select the anomaly type:
  - **Violence (1)**: Physical fights, assaults, aggression
  - **Theft (2)**: Stealing, pickpocketing, robbery
  - **Traffic Violation (3)**: Reckless driving, red light running
  - **Fire (4)**: Fire, smoke, explosions
  - **Vandalism (5)**: Property damage, destruction
  - **Trespassing (6)**: Unauthorized entry, fence climbing
  - **Normal (7)**: No anomaly (use for false alarms)

  ### Step 4: Rate Severity
  - **Low**: Minor, no immediate harm
  - **Medium**: Moderate risk, requires attention
  - **High**: Serious, requires immediate response
  - **Critical**: Life-threatening or extremely dangerous

  ### Tips:
  - Use slow motion for fast events
  - Pay attention to background activity, not just foreground
  - Mark the entire duration of the anomalous event
  - When in doubt between types, choose the most severe classification