DiDeMo Moment Retrieval

# DiDeMo Moment Retrieval Configuration
# Based on Hendricks et al., ICCV 2017
# Task: Localize text descriptions to video moments

annotation_task_name: "DiDeMo Moment Retrieval"
task_dir: "."

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

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

html_layout: |
  <div class="moment-retrieval">
    <div class="query-section" style="background: #e3f2fd; padding: 15px; border-radius: 8px; margin-bottom: 15px;">
      <h3 style="margin-top: 0;">🔍 Find this moment:</h3>
      <div class="query-text" style="font-size: 18px; font-weight: bold;">{{query}}</div>
    </div>
  </div>

annotation_schemes:
  - name: "moment_segment"
    description: |
      Mark the temporal segment that corresponds to the text description.
      Select the 5-second segment(s) that best match the query.
    annotation_type: "video_annotation"
    mode: "segment"
    labels:
      - name: "matches_query"
        color: "#22C55E"
        key_value: "m"
    zoom_enabled: true
    playback_rate_control: true
    frame_stepping: true
    show_timecode: true

  - name: "confidence"
    description: "How confident are you that this is the correct moment?"
    annotation_type: radio
    labels:
      - "Very confident - exact match"
      - "Confident - good match"
      - "Somewhat confident - partial match"
      - "Not confident - best guess"

  - name: "ambiguity"
    description: "Is the query ambiguous or could match multiple moments?"
    annotation_type: radio
    labels:
      - "Clear - only one possible match"
      - "Slightly ambiguous - 1-2 alternatives"
      - "Very ambiguous - multiple valid matches"

allow_all_users: true
instances_per_annotator: 60
annotation_per_instance: 3

annotation_instructions: |
  ## Video Moment Retrieval Task

  Given a text description, find the matching moment in the video.

  ### How to Annotate:
  1. Read the query carefully
  2. Watch the video to understand the content
  3. Mark the segment that BEST matches the description
  4. Rate your confidence

  ### Guidelines:
  - Videos are ~30 seconds, divided into 5-second segments
  - Select the segment(s) that match the query
  - If multiple segments match, select all of them
  - If no segment matches well, select the closest one

  ### Example Queries:
  - "The dog runs across the yard"
  - "Someone opens a door"
  - "A person laughs at something"

  ### Tips:
  - Focus on the action/event described
  - Consider synonyms (e.g., "runs" = "sprints")
  - The query may not describe every detail visible