FineSports Fine-grained Action Recognition

# FineSports Fine-grained Action Recognition Configuration
# Based on Xu et al., CVPR 2024
# Task: Track athletes and label fine-grained hierarchical actions in sports videos

annotation_task_name: "FineSports Fine-grained Action Recognition"
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:
  # Track athletes with bounding boxes
  - name: "athlete_tracking"
    description: |
      Draw bounding boxes around each athlete visible in the frame.
      Track the same athlete across frames using consistent IDs.
      Focus on athletes actively performing actions.
    annotation_type: "video_annotation"
    mode: "tracking"
    labels:
      - name: "athlete"
        color: "#3B82F6"
      - name: "referee"
        color: "#EF4444"
      - name: "coach"
        color: "#F59E0B"
    frame_stepping: true
    show_timecode: true
    playback_rate_control: true
    zoom_enabled: true
    video_fps: 30

  # Hierarchical action labels - Gymnastics
  - name: "gymnastics_actions"
    description: "Select all fine-grained gymnastics actions (if applicable)"
    annotation_type: multiselect
    labels:
      - "vault_run"
      - "vault_takeoff"
      - "vault_flight"
      - "vault_landing"
      - "balance_beam_mount"
      - "balance_beam_walk"
      - "balance_beam_leap"
      - "balance_beam_turn"
      - "balance_beam_dismount"
      - "floor_tumbling_pass"
      - "floor_dance_sequence"
      - "floor_leap_series"
      - "uneven_bars_swing"
      - "uneven_bars_release"
      - "uneven_bars_dismount"

  # Hierarchical action labels - Diving
  - name: "diving_actions"
    description: "Select all fine-grained diving actions (if applicable)"
    annotation_type: multiselect
    labels:
      - "approach"
      - "hurdle"
      - "takeoff_forward"
      - "takeoff_backward"
      - "takeoff_armstand"
      - "flight_tuck"
      - "flight_pike"
      - "flight_layout"
      - "flight_free"
      - "twist"
      - "somersault"
      - "entry"
      - "splash"

  # Hierarchical action labels - Figure Skating
  - name: "figure_skating_actions"
    description: "Select all fine-grained figure skating actions (if applicable)"
    annotation_type: multiselect
    labels:
      - "jump_axel"
      - "jump_lutz"
      - "jump_flip"
      - "jump_loop"
      - "jump_salchow"
      - "jump_toe_loop"
      - "spin_upright"
      - "spin_sit"
      - "spin_camel"
      - "spin_combination"
      - "step_sequence"
      - "spiral_sequence"
      - "lift"
      - "throw_jump"

  # Hierarchical action labels - General Sports
  - name: "general_sports_actions"
    description: "Select general sports actions applicable across sports"
    annotation_type: multiselect
    labels:
      - "preparation"
      - "execution"
      - "recovery"
      - "celebration"
      - "waiting"
      - "walking"
      - "running"
      - "jumping"
      - "landing"
      - "falling"

# User configuration
allow_all_users: true

# Task assignment
instances_per_annotator: 20
annotation_per_instance: 2

# Instructions
annotation_instructions: |
  ## FineSports Fine-grained Action Recognition Task

  Your goal is to track athletes and label their fine-grained actions in sports video clips.

  ### Step 1: Track Athletes
  - Draw bounding boxes around each athlete performing actions
  - Track the same athlete across frames with consistent IDs
  - Also mark referees and coaches if visible

  ### Step 2: Label Hierarchical Actions
  For each tracked athlete, select ALL fine-grained actions from the relevant sport category:

  **Gymnastics Actions:** vault, balance beam, floor, uneven bars phases
  **Diving Actions:** approach, takeoff, flight, twist, entry phases
  **Figure Skating Actions:** jumps, spins, step sequences, lifts
  **General Actions:** preparation, execution, recovery, etc.

  ### Important Notes:
  - Actions are organized hierarchically by sport type
  - One athlete can perform MULTIPLE actions simultaneously
  - Label the finest-grained action visible (e.g., "jump_axel" not just "jumping")
  - Only select actions from the sport category matching the video content
  - Use frame stepping for precise action boundary marking

  ### Tips:
  - Use playback speed control to slow down fast movements
  - Pay attention to transitions between action phases
  - Some actions are very brief (< 1 second), use frame stepping