The Activity Grouping extractor identifies human activities in videos, tracks them across frames, and groups similar activities for semantic video understanding.

​
Overview

The Activity Grouping feature extractor analyzes video content to detect human actions and activities, classifies them into categories, and groups similar activities throughout the video. This enables activity-based search and retrieval in video content.

​
Required Inputs

ParameterTypeRequiredDefaultDescription
video_urlstringYes-URL pointing to video file. Supported formats: MP4, MOV, AVI
min_activity_durationfloatNo1.0Minimum activity duration in seconds
detection_thresholdfloatNo0.7Confidence threshold for activity detection (0.0-1.0)
similarity_thresholdfloatNo0.8Similarity threshold for grouping activities (0.0-1.0)

​
Configurations

​
Detection Modes

The extractor supports different detection modes based on use case requirements:

ModeDescriptionBest For
standardBalanced accuracy and performanceGeneral activity detection
high-accuracyPrioritizes detection accuracyAction recognition
high-performancePrioritizes processing speedReal-time applications

​
Configuration Examples

Standard
{
  "detection_mode": "standard",
  "min_activity_duration": 1.0,
  "detection_threshold": 0.7,
  "similarity_threshold": 0.8
}
High Accuracy
{
  "detection_mode": "high-accuracy",
  "min_activity_duration": 0.5,
  "detection_threshold": 0.85,
  "similarity_threshold": 0.9,
  "detect_fine_grained": true
}
High Performance
{
  "detection_mode": "high-performance",
  "min_activity_duration": 1.5,
  "detection_threshold": 0.6,
  "similarity_threshold": 0.7,
  "frame_sampling": 5
}

​
Activity Grouping Options

OptionTypeDefaultDescription
detect_fine_grainedbooleanfalseDetect detailed sub-categories of activities
track_participantsbooleantrueTrack the people participating in activities
generate_embeddingbooleantrueGenerate activity embedding for similarity matching
extract_representative_framesbooleantrueExtract key frames representing the activity
frame_samplinginteger1Process every Nth frame for performance

​
Configuration Examples

Sample
{
  "detect_fine_grained": true,             // Detect detailed activity types
  "track_participants": true,              // Track people in activities
  "generate_embedding": true,              // Generate activity embedding
  "extract_representative_frames": true,   // Extract key frames
  "frame_sampling": 2                      // Process every 2nd frame
}

​
Processing Flow

​
Output Schema

This feature extractor will output as features in the feature store.

{
  "document_id": "doc_abc123",
  "collection_id": "col_xyz789",
  "source_object_id": "obj_def456",
  
  // Activity grouping results
  "activity": {
    "id": "activity_12345",
    "class": "dancing",
    "fine_grained_class": "ballroom_dancing",
    "confidence": 0.92,
    "instances": [
      {
        "start_frame": 120,
        "end_frame": 180,
        "start_time": 4.0,
        "end_time": 6.0,
        "duration": 2.0,
        "participants": [
          "person_123", "person_456"
        ],
        "confidence": 0.94,
        "bounding_box": {
          "x": 125,
          "y": 80,
          "width": 320,
          "height": 450
        }
      },
      {
        "start_frame": 300,
        "end_frame": 390,
        "start_time": 10.0,
        "end_time": 13.0,
        "duration": 3.0,
        "participants": [
          "person_123", "person_456"
        ],
        "confidence": 0.91,
        "bounding_box": {
          "x": 145,
          "y": 85,
          "width": 310,
          "height": 440
        }
      }
      // Additional instances...
    ],
    "embedding": {
      "model": "activity_encoder_v1",
      "dimension": 512,
      "vector": [0.12, 0.34, ...], // Truncated for brevity
      "normalized": true
    },
    "total_duration": 5.0,              // Total duration of all instances in seconds
    "representative_frame_urls": [
      "https://storage.example.com/activities/activity_12345_1.jpg",
      "https://storage.example.com/activities/activity_12345_2.jpg"
    ]
  },
  
  // Video metadata
  "video": {
    "filename": "dance_performance.mp4",
    "width": 1920,
    "height": 1080,
    "fps": 30,
    "duration": 180.5
  }
}

Was this page helpful?

Suggest editsRaise issue