Charades-STA Temporal Grounding

# Charades-STA Temporal Grounding Configuration
# Based on Gao et al., ICCV 2017
# Task: Ground language descriptions to video segments

annotation_task_name: "Charades-STA Temporal Grounding"
task_dir: "."

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

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

annotation_schemes:
  - name: "grounded_segment"
    description: |
      Mark the EXACT temporal segment where the described action occurs.
      Read the query carefully and find where it happens in the video.
    annotation_type: "video_annotation"
    mode: "segment"
    labels:
      - name: "grounded_moment"
        color: "#22C55E"
        key_value: "g"
    frame_stepping: true
    show_timecode: true
    playback_rate_control: true
    video_fps: 24

  - name: "grounding_confidence"
    description: "How confident are you in the temporal boundaries?"
    annotation_type: radio
    labels:
      - "Very confident - clear start and end"
      - "Confident - boundaries are reasonably clear"
      - "Somewhat confident - boundaries are approximate"
      - "Not confident - hard to determine exact timing"

  - name: "query_ambiguity"
    description: "Is the language query ambiguous?"
    annotation_type: radio
    labels:
      - "Clear - unambiguous description"
      - "Slightly ambiguous - minor interpretation needed"
      - "Ambiguous - multiple interpretations possible"
      - "Very ambiguous - unclear what to look for"

  - name: "action_visible"
    description: "Is the described action visible in the video?"
    annotation_type: radio
    labels:
      - "Fully visible - entire action shown"
      - "Partially visible - action partly shown"
      - "Barely visible - hard to see"
      - "Not visible - action not in video"

allow_all_users: true
instances_per_annotator: 60
annotation_per_instance: 2

annotation_instructions: |
  ## Charades-STA Temporal Grounding

  Ground natural language descriptions to video segments.

  ### Task:
  Given a sentence describing an action, mark the EXACT segment where that action occurs.

  ### Example:
  - Query: "Person opens a door"
  - Your task: Find and mark the segment where a person opens a door

  ### Guidelines:
  - Read the query BEFORE watching the video
  - Mark from action START to action END
  - Include preparation if it's part of the action
  - The segment should be as tight as possible

  ### Boundary Rules:
  - START: When the person begins the action (not before)
  - END: When the action is complete (not after)
  - If the action repeats, mark only the FIRST occurrence

  ### Common issues:
  - Actions may be brief (1-2 seconds)
  - Multiple similar actions may occur
  - Some queries may not have a match (mark "not visible")

  ### Tips:
  - Use frame stepping for precise boundaries
  - The action should match the ENTIRE query, not just part
  - Pay attention to object mentions ("opens a door" vs "opens a window")