This weblog is co-authored by James Le, Head of Developer Expertise – TwelveLabs

The exponential development of video content material has created each alternatives and challenges. Content material creators, entrepreneurs, and researchers at the moment are confronted with the daunting activity of effectively looking, analyzing, and extracting priceless insights from huge video libraries. Conventional search strategies akin to keyword-based textual content search typically fall quick when coping with video content material to investigate the visible content material, spoken phrases, or contextual parts inside the video itself, leaving organizations struggling to successfully search via and unlock the total potential of their multimedia belongings.

With the combination of TwelveLabs’ Embed API and Amazon OpenSearch Service, we are able to work together with and derive worth from video content material. By utilizing TwelveLabs‘ superior AI-powered video understanding know-how and OpenSearch Service’s search and analytics capabilities, we are able to now carry out superior video discovery and acquire deeper insights.

On this weblog publish, we present you the method of integrating TwelveLabs Embed API with OpenSearch Service to create a multimodal search resolution. You’ll learn to generate wealthy, contextual embeddings from video content material and use OpenSearch Service’s vector database capabilities to allow search functionalities. By the top of this publish, you’ll be outfitted with the data to implement a system that may rework the best way your group handles and extracts worth from video content material.

TwelveLabs’ multimodal embeddings course of visible, audio, and textual content alerts collectively to create unified representations, capturing the direct relationships between these modalities. This unified method delivers exact, context-aware video search that matches human understanding of video content material. Whether or not you’re a developer trying to improve your purposes with superior video search capabilities, or a enterprise chief in search of to optimize your content material administration methods, this publish will offer you the instruments and steps to implement multimodal seek for your organizational information.

About TwelveLabs

TwelveLabs is an Superior AWS Accomplice and AWS Market Vendor that provides video understanding options. Embed API is designed to revolutionize the way you work together with and extract worth from video content material.

At its core, the Embed API transforms uncooked video content material into significant, searchable information through the use of state-of-the-art machine studying fashions. These fashions extract and characterize advanced video data within the type of dense vector embeddings, every a typical 1024-dimensional vector that captures the essence of the video content material throughout a number of modalities (picture, textual content, and audio).

Key options of TwelveLabs Embed API

Under are the important thing options of TwelveLabs Embed API:

• Multimodal understanding: The API generates embeddings that encapsulate numerous features of the video, together with visible expressions, physique language, spoken phrases, and general context.
• Temporal coherence: In contrast to static image-based fashions, TwelveLabs’ embeddings seize the interrelations between completely different modalities over time, offering a extra correct illustration of video content material.
• Flexibility: The API helps native processing of all modalities current in movies, eliminating the necessity for separate text-only or image-only fashions.
• Excessive efficiency: By utilizing a video-native method, the Embed API offers extra correct and temporally coherent interpretation of video content material in comparison with conventional CLIP-like fashions.

Advantages and use circumstances

The Embed API presents quite a few benefits for builders and companies working with video content material:

• Enhanced Search Capabilities: Allow highly effective multimodal search throughout video libraries, permitting customers to search out related content material based mostly on visible, audio, or textual queries.
• Content material Suggestion: Enhance content material advice methods by understanding the deep contextual similarities between movies.
• Scene Detection and Segmentation: Robotically detect and phase completely different scenes inside movies for simpler navigation and evaluation.
• Content material Moderation: Effectively establish and flag inappropriate content material throughout massive video datasets.

Use circumstances embody:

• Anomaly detection
• Variety sorting
• Sentiment evaluation
• Suggestions

Structure overview

The structure for utilizing TwelveLabs Embed API and OpenSearch Service for superior video search consists of the next parts:

• TwelveLabs Embed API: This API generates 1024-dimensional vector embeddings from video content material, capturing visible, audio, and textual parts.
• OpenSearch Vector Database: Shops and indexes the video embeddings generated by TwelveLabs.
• Secrets and techniques Supervisor to retailer secrets and techniques akin to API entry keys, and the Amazon OpenSearch Service username and password.
• Integration of TwelveLabs SDK and the OpenSearch Service shopper to course of movies, generate embeddings, and index them in OpenSearch Service.

The next diagram illustrates:

1. A video file is saved in Amazon Easy Storage Service (Amazon S3). Embeddings of the video file are created utilizing TwelveLabs Embed API.
2. Embeddings generated from the TwelveLabs Embed API at the moment are ingested to Amazon OpenSearch Service.
3. Customers can search the video embeddings utilizing textual content, audio, or picture. The person makes use of TwelveLabs Embed API to create the corresponding embeddings.
4. The person searches video embeddings in Amazon OpenSearch Service and retrieves the corresponding vector.

The use case

For the demo, you’ll work on these movies: Robin chicken forest Video by Federico Maderno from Pixabay and Island Video by Bellergy RC from Pixabay.

Nonetheless, the use case could be expanded to varied different segments. For instance, the information group struggles with:

1. Needle-in-haystack searches via 1000’s of hours of archival footage
2. Guide metadata tagging that misses nuanced visible and audio context
3. Cross-modal queries akin to querying a video assortment utilizing textual content or audio descriptions
4. Speedy content material retrieval for breaking information tie-ins

By integrating TwelveLabs Embed API with OpenSearch Service, you’ll be able to:

• Generate 1024-dimensional embeddings capturing every video’s visible ideas. The embeddings are additionally able to extracting spoken narration, on-screen textual content, and audio cues.
• Allow multimodal search capabilities permitting customers to:
  • Discover particular demonstrations utilizing text-based queries.
  • Find actions via image-based queries.
  • Establish segments utilizing audio sample matching.
• Scale back search time from hours to seconds for advanced queries.

Resolution walkthrough

GitHub repository comprises a pocket book with detailed walkthrough directions for implementing superior video search capabilities by combining TwelveLabs’ Embed API with Amazon OpenSearch Service.

Stipulations

Earlier than you proceed additional, confirm that the next conditions are met:

• Affirm that you’ve an AWS account. Register to the AWS account.
• Create a TwelveLabs account as a result of will probably be required to get the API Key. TwelveLabs supply free tier pricing however you’ll be able to improve if vital to fulfill your requirement.
• Have an Amazon OpenSearch Service area. For those who don’t have an current area, you’ll be able to create one utilizing the steps outlined in our public documentation for Creating and Managing Amazon OpenSearch Service Area. Ensure that the OpenSearch Service area is accessible out of your Python setting. You may also use Amazon OpenSearch Serverless for this use case and replace the interactions to OpenSearch Serverless utilizing AWS SDKs.

Step 1: Arrange the TwelveLabs SDK

Begin by establishing the TwelveLabs SDK in your Python setting:

1. Receive your API key from TwelveLabs Dashboard.
2. Observe steps right here to create a secret in AWS Secrets and techniques Supervisor. For instance, identify the key as TL_API_Key.Word down the ARN or identify of the key (TL_API_Key) to retrieve. To retrieve a secret from one other account, you have to use an ARN.For an ARN, we advocate that you simply specify a whole ARN relatively than a partial ARN. See Discovering a secret from a partial ARN.Use this worth for the SecretId within the code block beneath.

import boto3
import json
secrets_manager_client=boto3.shopper("secretsmanager")
API_secret=secrets_manager_client.get_secret_value(
SecretId="TL_API_KEY"
)
TL_API_KEY=json.hundreds(API_secret["SecretString"])["TL_API_Key"]

Step 2: Generate video embeddings

Use the Embed API to create multimodal embeddings which are contextual vector representations on your movies and texts. TwelveLabs video embeddings seize all of the refined cues and interactions between completely different modalities, together with the visible expressions, physique language, spoken phrases, and the general context of the video, encapsulating the essence of all these modalities and their interrelations over time.

To create video embeddings, you have to first add your movies, and the platform should end processing them. Importing and processing movies require a while. Consequently, creating embeddings is an asynchronous course of comprised of three steps:

1. Add and course of a video: While you begin importing a video, the platform creates a video embedding activity and returns its distinctive activity identifier.
2. Monitor the standing of your video embedding activity: Use the distinctive identifier of your activity to verify its standing periodically till it’s accomplished.
3. Retrieve the embeddings: After the video embedding activity is accomplished, retrieve the video embeddings by offering the duty identifier. Be taught extra within the docs.

Video processing implementation

This demo relies upon upon some video information. To make use of this, you’ll obtain two mp4 recordsdata and add it to an Amazon S3 bucket.

1. Click on on the hyperlinks containing the Robin chicken forest Video by Federico Maderno from Pixabay and Island Video by Bellergy RC from Pixabay movies.
2. Obtain the 21723-320725678_small.mp4 and 2946-164933125_small.mp4 recordsdata.
3. Create an S3 bucket in case you don’t have one already. Observe the steps within the Making a bucket doc. Word down the identify of the bucket and change it the code block beneath (Eg., MYS3BUCKET).
4. Add the 21723-320725678_small.mp4 and 2946-164933125_small.mp4 video recordsdata to the S3 bucket created within the step above by following the steps within the Importing objects doc. Word down the identify of the objects and change it the code block beneath (Eg., 21723-320725678_small.mp4 and 2946-164933125_small.mp4)

s3_client=boto3.shopper("s3")
bird_video_data=s3_client.download_file(Bucket="MYS3BUCKET",  Key='21723-320725678_small.mp4', Filename="robin-bird.mp4")
island_video_data=s3_client.download_file(Bucket="MYS3BUCKET",  Key='2946-164933125_small.mp4', Filename="island.mp4")

def print_segments(segments: Record[SegmentEmbedding], max_elements: int = 1024):
    for phase in segments:
        print(
            f"  embedding_scope={phase.embedding_scope} start_offset_sec={phase.start_offset_sec} end_offset_sec={phase.end_offset_sec}"
        )
        print(f"  embeddings: {phase.embeddings_float[:max_elements]}")

# Initialize shopper with API key
twelvelabs_client = TwelveLabs(api_key=TL_API_KEY)

video_files=["robin-bird.mp4", "island.mp4"]
duties=[]

Embedding technology course of

With the SDK configured, generate embeddings on your video and monitor activity completion with real-time updates. Right here you utilize the Marengo 2.7 mannequin to generate the embeddings:

for video in video_files:
    # Create embedding activity
    activity = twelvelabs_client.embed.activity.create(
        model_name="Marengo-retrieval-2.7",
        video_file=video
    )
    print(
        f"Created activity: id={activity.id} engine_name={activity.model_name} standing={activity.standing}"
    )
    
    def on_task_update(activity: EmbeddingsTask):
        print(f"  Standing={activity.standing}")
    
    standing = activity.wait_for_done(
        sleep_interval=2,
        callback=on_task_update
    )
    print(f"Embedding achieved: {standing}")
    
    # Retrieve and examine outcomes
    activity = activity.retrieve()
    if activity.video_embedding shouldn't be None and activity.video_embedding.segments shouldn't be None:
        print_segments(activity.video_embedding.segments)
    duties.append(activity)

Key options demonstrated embody:

• Multimodal seize: 1024-dimensional vectors encoding visible, audio, and textual options
• Mannequin specificity: Utilizing Marengo-retrieval-2.7 optimized for scientific content material
• Progress monitoring: Actual-time standing updates throughout embedding technology

Anticipated output

Created activity: id=67ca93a989d8a564e80dc3ba engine_name=Marengo-retrieval-2.7 standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=processing
  Standing=prepared
Embedding achieved: prepared
  embedding_scope=clip start_offset_sec=0.0 end_offset_sec=6.0
  embeddings: [0.022429451, 0.00040668788, -0.01825908, -0.005862708, -0.03371106, 
-6.357456e-05, -0.015320076, -0.042556215, -0.02782445, -0.00019097517, 0.03258314, 
-0.0061399476, -0.00049206393, 0.035632476, 0.028209884, 0.02875258, -0.035486065, 
-0.11288028, -0.040782217, -0.0359422, 0.015908664, -0.021092793, 0.016303983, 
0.06351931,…………………

Step 3: Set up OpenSearch

To enable vector search capabilities, you first need to set up an OpenSearch client and test the connection. Follow these steps:

Install the required libraries

Install the necessary Python packages for working with OpenSearch:

!pip install opensearch-py
!pip install botocore
!pip install requests-aws4auth

Configure the OpenSearch client

Set up the OpenSearch client with your host details and authentication credentials:

from opensearchpy import OpenSearch, RequestsHttpConnection, helpers
from requests_aws4auth import AWS4Auth
from requests.auth import HTTPBasicAuth

# OpenSearch connection configuration
# host="your-host.aos.us-east-1.on.aws"
host="search-new-domain-mbgs7wth6r5w6hwmjofntiqcge.aos.us-east-1.on.aws"
port = 443  # Default HTTPS port

# Get OpenSearch username secret from Secrets Manager
opensearch_username=secrets_manager_client.get_secret_value(
    SecretId="AOS_username"
)
opensearch_username_string=json.loads(opensearch_username["SecretString"])["AOS_username"]

# Get OpenSearch password secret from Secrets and techniques Supervisor
opensearch_password = secrets_manager_client.get_secret_value(
    SecretId="AOS_password"
)
opensearch_password_string=json.hundreds(opensearch_password["SecretString"])["AOS_password"]

auth=(opensearch_username_string, opensearch_password_string)

# Create the shopper configuration
client_aos = OpenSearch(
    hosts=[{'host': host, 'port': port}],
    http_auth=auth,
    use_ssl=True,
    verify_certs=True,
    connection_class=RequestsHttpConnection
)

# Check the connection
strive:
    # Get cluster data
    cluster_info = client_aos.information()
    print("Efficiently related to OpenSearch")
    print(f"Cluster information: {cluster_info}")
besides Exception as e:
    print(f"Connection failed: {str(e)}")

Anticipated output

If the connection is profitable, it’s best to see a message like the next:

Efficiently related to OpenSearch
Cluster information: {'identify': 'bb36e8d98ee7bd517891ecd714bfb9d7', ...}

This confirms that your OpenSearch shopper is correctly configured and prepared to be used.

Step 4: Create an index in OpenSearch Service

Subsequent, you create an index optimized for vector search to retailer the embeddings generated by the TwelveLabs Embed API.

Outline the index configuration

The index is configured to help k-nearest neighbor (kNN) search with a 1024-dimensional vector subject. You’ll these values for this demo however comply with these finest practices to search out acceptable values on your utility. Right here’s the code:

# Outline the improved index configuration
index_name="twelvelabs_index"
new_vector_index_definition = {
    "settings": {
        "index": {
            "knn": "true",
            "number_of_shards": 1,
            "number_of_replicas": 0
        }
    },
    "mappings": {
        "properties": {
            "embedding_field": {
                "kind": "knn_vector",
                "dimension": 1024
            },
            "video_title": {
                "kind": "textual content",
                "fields": {
                    "key phrase": {
                        "kind": "key phrase"
                    }
                }
            },
            "segment_start": {
                "kind": "date"
            },
            "segment_end": {
                "kind": "date"
            },
            "segment_id": {
                "kind": "textual content"
            }
        }
    }
}

Create the Index

Use the next code to create the index in OpenSearch Service:

# Create the index in OpenSearch
response = client_aos.indices.create(index=index_name, physique=new_vector_index_definition, ignore=400)

# Retrieve and show index particulars to substantiate creation
index_info = client_aos.indices.get(index=index_name)
print(index_info)

Anticipated output

After working this code, it’s best to see particulars of the newly created index. For instance:

{'twelvelabs_index': {'aliases': {}, 'mappings': {'properties': {'embedding_field': {'kind': 'knn_vector', 'dimension': 1024}}}, 'settings': {...}}}

The next screenshot confirms that an index named twelvelabs_index has been efficiently created with a knn_vector subject of dimension 1024 and different specified settings. With these steps accomplished, you now have an operational OpenSearch Service area configured for vector search. This index will function the repository for storing embeddings generated from video content material, enabling superior multimodal search capabilities.

Step 5: Ingest embeddings to the created index in OpenSearch Service

With the TwelveLabs Embed API efficiently producing video embeddings and the OpenSearch Service index configured, the following step is to ingest these embeddings into the index. This course of helps be sure that the embeddings are saved in OpenSearch Service and made searchable for multimodal queries.

Embedding ingestion course of

The next code demonstrates course of and index the embeddings into OpenSearch Service:

from opensearchpy.helpers import bulk

def generate_actions(duties, video_files):
    rely = 0
    for activity in duties:
        # Examine if video embeddings can be found
        if activity.video_embedding shouldn't be None and activity.video_embedding.segments shouldn't be None:
            embeddings_doc = activity.video_embedding.segments
            
            # Generate actions for bulk indexing
            for doc_id, elt in enumerate(embeddings_doc):
                yield {
                    '_index': index_name,
                    '_id': doc_id,
                    '_source': {
                        'embedding_field': elt.embeddings_float,
                        'video_title': video_files[count],
                        'segment_start': elt.start_offset_sec,
                        'segment_end': elt.end_offset_sec,
                        'segment_id': doc_id
                    }
                }
        print(f"Ready bulk indexing information for activity {rely}")
        rely += 1

# Carry out bulk indexing
strive:
    success, failed = bulk(client_aos, generate_actions(duties, video_files))
    print(f"Efficiently listed {success} paperwork")
    if failed:
        print(f"Didn't index {len(failed)} paperwork")
besides Exception as e:
    print(f"Error throughout bulk indexing: {e}")

Rationalization of the code

1. Embedding extraction: The video_embedding.segments object comprises an inventory of phase embeddings generated by the TwelveLabs Embed API. Every phase represents a particular portion of the video.
2. Doc creation: For every phase, a doc is created with a key (embedding_field) that shops its 1024-dimensional vector, video_title with the title of the video, segment_start and segment_end indicating the timestamp of the video phase, and a segment_id.
3. Indexing in OpenSearch: The index() technique uploads every doc to the twelvelabs_index created earlier. Every doc is assigned a novel ID (doc_id) based mostly on its place within the listing.

Anticipated output

After the script runs efficiently, you will note:

• A printed listing of embeddings being listed.
• A affirmation message:

Ready bulk indexing information for activity 0
Ready bulk indexing information for activity 1
Efficiently listed 6 paperwork

End result

At this stage, all video phase embeddings at the moment are saved in OpenSearch and prepared for superior multimodal search operations, akin to text-to-video or image-to-video queries. This units up the muse for performing environment friendly and scalable searches throughout your video content material.

Step 6: Carry out vector search in OpenSearch Service

After embeddings are generated, you utilize it as a question vector to carry out a kNN search within the OpenSearch Service index. Under are the capabilities to carry out vector search and format the search outcomes:

# Perform to carry out vector search
def search_similar_segments(query_vector, ok=5):
    question = {
        "measurement": ok,
        "_source": ["video_title", "segment_start", "segment_end", "segment_id"],
        "question": {
            "knn": {
                "embedding_field": {
                    "vector": query_vector,
                    "ok": ok
                }
            }
        }
    }
    
    response = client_aos.search(
        index=index_name,
        physique=question
    )

    outcomes = []
    for hit in response['hits']['hits']:
        consequence = {
            'rating': hit['_score'],
            'title': hit['_source']['video_title'],
            'start_time': hit['_source']['segment_start'],
            'end_time': hit['_source']['segment_end'],
            'segment_id': hit['_source']['segment_id']
        }
        outcomes.append(consequence)

    return (outcomes)

# Perform to format search outcomes
def print_search_results(outcomes):
    print("nSearch Outcomes:")
    print("-" * 50)
    for i, lead to enumerate(outcomes, 1):
        print(f"nResult {i}:")
        print(f"Video: {consequence['title']}")
        print(f"Time Vary: {consequence['start_time']} - {consequence['end_time']}")
        print(f"Similarity Rating: {consequence['score']:.4f}")

Key factors:

• The _source subject comprises the video title, phase begin, phase finish, and phase id comparable to the video embeddings.
• The embedding_field within the question corresponds to the sphere the place video embeddings are saved.
• The ok parameter specifies what number of high outcomes to retrieve based mostly on similarity.

Step 7:Performing text-to-video search

You should use text-to-video search to retrieve video segments which are most related to a given textual question. On this resolution, you’ll do that through the use of TwelveLabs’ textual content embedding capabilities and OpenSearch’s vector search performance. Right here’s how one can implement this step:

Generate textual content embeddings

To carry out a search, you first have to convert the textual content question right into a vector illustration utilizing the TwelveLabs Embed API:

from typing import Record
from twelvelabs.fashions.embed import SegmentEmbedding

def print_segments(segments: Record[SegmentEmbedding], max_elements: int = 1024):
    for phase in segments:
        print(
            f"  embedding_scope={phase.embedding_scope} start_offset_sec={phase.start_offset_sec} end_offset_sec={phase.end_offset_sec}"
        )
        print(f"  embeddings: {phase.embeddings_float[:max_elements]}")

# Create textual content embeddings for the question
text_res = twelvelabs_client.embed.create(
    model_name="Marengo-retrieval-2.7",
    textual content="Chook consuming meals",  # Change along with your desired question
)

print("Created a textual content embedding")
print(f" Mannequin: {text_res.model_name}")

# Extract and examine the generated textual content embeddings
if text_res.text_embedding shouldn't be None and text_res.text_embedding.segments shouldn't be None:
    print_segments(text_res.text_embedding.segments)

vector_search = text_res.text_embedding.segments[0].embeddings_float
print("Generated Textual content Embedding Vector:", vector_search)

Key factors:

• The Marengo-retrieval-2.7 mannequin is used to generate a dense vector embedding for the question.
• The embedding captures the semantic that means of the enter textual content, enabling efficient matching with video embeddings.

Carry out vector search in OpenSearch Service

After the textual content embedding is generated, you utilize it as a question vector to carry out a kNN search within the OpenSearch index:

# Outline the vector search question
query_vector = vector_search
text_to_video_search = search_similar_segments(query_vector)
# print(text_video_search)
print_search_results(text_to_video_search)

Anticipated output

The next illustrates comparable outcomes retrieved from OpenSearch.

Search Outcomes:
--------------------------------------------------

End result 1:
Video: robin-bird.mp4
Time Vary: 18.0 - 21.087732
Similarity Rating: 0.4409

End result 2:
Video: robin-bird.mp4
Time Vary: 12.0 - 18.0
Similarity Rating: 0.4300

End result 3:
Video: island.mp4
Time Vary: 0.0 - 6.0
Similarity Rating: 0.3624

Insights from outcomes

• Every consequence features a similarity rating indicating how carefully it matches the question, a time vary indicating the beginning and finish offset in seconds, and the video title.
• Observe that the highest 2 outcomes correspond to the robin chicken video segments matching the Chook consuming meals question.

This course of demonstrates how textual queries akin to Chook consuming meals can successfully retrieve related video segments from an listed library utilizing TwelveLabs’ multimodal embeddings and OpenSearch’s highly effective vector search capabilities.

Step 8: Carry out audio-to-video search

You should use audio-to-video search to retrieve video segments which are most related to a given audio enter. By utilizing TwelveLabs’ audio embedding capabilities and OpenSearch’s vector search performance, you’ll be able to match audio options with video embeddings within the index. Right here’s implement this step:

Generate audio embeddings

To carry out the search, you first convert the audio enter right into a vector illustration utilizing the TwelveLabs Embed API:

# Create audio embeddings for the enter audio file
audio_res = twelvelabs_client.embed.create(
    model_name="Marengo-retrieval-2.7",
    audio_file="audio-data.mp3",  # Change along with your desired audio file
)

# Print particulars of the generated embedding
print(f"Created audio embedding: model_name={audio_res.model_name}")
print(f" Mannequin: {audio_res.model_name}")

# Extract and examine the generated audio embeddings
if audio_res.audio_embedding shouldn't be None and audio_res.audio_embedding.segments shouldn't be None:
    print_segments(audio_res.audio_embedding.segments)

# Retailer the embedding vector for search
vector_search = audio_res.audio_embedding.segments[0].embeddings_float
print("Generated Audio Embedding Vector:", vector_search)

Key factors:

• The Marengo-retrieval-2.7 mannequin is used to generate a dense vector embedding for the enter audio.
• The embedding captures the semantic options of the audio, akin to rhythm, tone, and patterns, enabling efficient matching with video embeddings

Carry out vector search in OpenSearch Service

After the audio embedding is generated, you utilize it as a question vector to carry out a k-nearest neighbor (kNN) search in OpenSearch:

# Carry out vector search
query_vector = vector_search
audio_to_video_search = search_similar_segments(query_vector)
# print(text_video_search)
    
print_search_results(audio_to_video_search)

Anticipated output

The next reveals video segments retrieved from OpenSearch Service based mostly on their similarity to the enter audio.

Search Outcomes:
--------------------------------------------------

End result 1:
Video: island.mp4
Time Vary: 6.0 - 12.0
Similarity Rating: 0.2855

End result 2:
Video: robin-bird.mp4
Time Vary: 18.0 - 21.087732
Similarity Rating: 0.2841

End result 3:
Video: robin-bird.mp4
Time Vary: 12.0 - 18.0
Similarity Rating: 0.2837

End result 4:
Video: island.mp4
Time Vary: 0.0 - 6.0
Similarity Rating: 0.2835

Right here discover that segments from each movies are returned with a low similarity rating.

Step 9: Performing images-to-video search

You should use image-to-video search to retrieve video segments which are visually just like a given picture. By utilizing TwelveLabs’ picture embedding capabilities and OpenSearch Service’s vector search performance, you’ll be able to match visible options from a picture with video embeddings within the index. Right here’s implement this step:

Generate Picture Embeddings

To carry out the search, you first convert the enter picture right into a vector illustration utilizing the TwelveLabs Embed API:

# Create picture embeddings for the enter picture file
image_res = twelvelabs_client.embed.create(
    model_name="Marengo-retrieval-2.7",
    image_file="image-data.jpg",  # Change along with your desired picture file
)

# Print particulars of the generated embedding
print(f"Created picture embedding: model_name={image_res.model_name}")
print(f" Mannequin: {image_res.model_name}")

# Extract and examine the generated picture embeddings
if image_res.image_embedding shouldn't be None and image_res.image_embedding.segments shouldn't be None:
    print_segments(image_res.image_embedding.segments)

# Retailer the embedding vector for search
vector_search = image_res.image_embedding.segments[0].embeddings_float
print("Generated Picture Embedding Vector:", vector_search)

Key factors:

• The Marengo-retrieval-2.7 mannequin is used to generate a dense vector embedding for the enter picture.
• The embedding captures visible options akin to shapes, colours, and patterns, enabling efficient matching with video embeddings

Carry out vector search in OpenSearch

After the picture embedding is generated, you utilize it as a question vector to carry out a k-nearest neighbor (kNN) search in OpenSearch:

# Carry out vector search
query_vector = vector_search
image_to_video_search = search_similar_segments(query_vector)
# print(text_video_search)
    
print_search_results(image_to_video_search)

Anticipated output

The next reveals video segments retrieved from OpenSearch based mostly on their similarity to the enter picture.

Search Outcomes:
--------------------------------------------------

End result 1:
Video: island.mp4
Time Vary: 6.0 - 12.0
Similarity Rating: 0.5616

End result 2:
Video: island.mp4
Time Vary: 0.0 - 6.0
Similarity Rating: 0.5576

End result 3:
Video: robin-bird.mp4
Time Vary: 12.0 - 18.0
Similarity Rating: 0.4592

End result 4:
Video: robin-bird.mp4
Time Vary: 18.0 - 21.087732
Similarity Rating: 0.4540

Observe that picture of an ocean was used to go looking the movies. Video clips from the island video are retrieved with a better similarity rating within the first 2 outcomes.

Clear up

To keep away from costs, delete assets created whereas following this publish. For Amazon OpenSearch Service domains, navigate to the AWS Administration Console for Amazon OpenSearch Service dashboard and delete the area.

Conclusion

The combination of TwelveLabs Embed API with OpenSearch Service offers a cutting-edge resolution for superior video search and evaluation, unlocking new prospects for content material discovery and insights. By utilizing TwelveLabs’ multimodal embeddings, which seize the intricate interaction of visible, audio, and textual parts in movies, and mixing them with OpenSearch Service’s sturdy vector search capabilities, this resolution permits extremely nuanced and contextually related video search.

As industries more and more depend on video content material for communication, training, advertising, and analysis, this superior search resolution turns into indispensable. It empowers companies to extract hidden insights from their video content material, improve person experiences in video-centric purposes and make data-driven choices based mostly on complete video evaluation

This integration not solely addresses present challenges in managing video content material but additionally lays the muse for future improvements in how we work together with and derive worth from video information.

Get began

Able to discover the facility of TwelveLabs Embed API? Begin your free trial at this time by visiting TwelveLabs Playground to enroll and obtain your API key.

For builders trying to implement this resolution, comply with our detailed step-by-step information on GitHub to combine TwelveLabs Embed API with OpenSearch Service and construct your individual superior video search utility.

Unlock the total potential of your video content material at this time!

In regards to the Authors

James Le runs the Developer Expertise operate at TwelveLabs. He works with companions, builders, and researchers to convey state-of-the-art video basis fashions to varied multimodal video understanding use circumstances.

Gitika is an Senior WW Information & AI Accomplice Options Architect at Amazon Net Companies (AWS). She works with companions on technical initiatives, offering architectural steering and enablement to construct their analytics apply.

Kruthi is a Senior Accomplice Options Architect specializing in AI and ML. She offers technical steering to AWS Companions in following finest practices to construct safe, resilient, and extremely obtainable options within the AWS Cloud.