Hikvision Network Video Recorders

How AcuSeek works

Based on a Large Multimodal Model (LLM), images and their corresponding natural language descriptions are mapped to a shared high-dimensional vector space. This enables versatile functionalities, such as open-ended text-to-image search, image-to-image search, and alarm triggering based on the presence of persons, vehicles, objects, events, or locations.
Massive image-text pairs
"Dog in the snow"
Image Features: Outline, Texture, Color...(Unstructured characteristics)
Image Encoder
Text Encoder
Text Features: Word Frequency, Order, Grammar.-(statistical characteristics)
Large Multimodal Model
AcuSeek: A man in a black top, white pants, and glasses is walking his dog.
Training

Massive image-text pairs
"Dog in the snow"
Image Features: Outline, Texture, Color... (Unstructured characteristics)
Image Encoder
Text Encoder
Text Features: Word Frequency, Order, Grammar... (statistical characteristics)
Large Multimodal Model
AcuSeek:

A man in a black top, white pants, and glasses is walking his dog.
Based on a Large Multimodal Model (LLM), images and their corresponding natural language descriptions are mapped to a shared high-dimensional vector space. This enables versatile functionalities, such as open-ended text-to-image search, image-to-image search, and alarm triggering based on the presence of persons, vehicles, objects, events, or locations.
Training
How AcuSeek works

Search by text

Search by text

AcuSeek NVR supports natural-language-based video/image retrieval applications, allowing second-level retrieval on target images by inputting phrases or words.

Search Broadly: Supports both high-frequency object searches in security scenarios, such as cars and non-motorized vehicles, and feature-based searches for specific objects or unusual activities.
Search Instantly: Provides second-level search results.
Search Accurately: Extracts features from images and aligns them with key semantic features from the search language to achieve high-precision search results.

Bridging text queries and visual matching

Intelligent video search

Step 1:

AcuSeek identifies objects associated with a specific feature (e.g., an Audi car) or another object (e.g., a black backpack) by analyzing video footage and user-defined criteria such as time and location.
Step 2:

AcuSearch then pinpoints a specific object (e.g., a grey Audi SUV) across all scenes, providing a timeline of its appearances and related footage for further check.
Describe by text
Retrieved objects
Click AcuSeek to view

Step 1: AcuSeek identifies objects associated with a specific feature (e.g., an Audi car) or another object (e.g., a black backpack) by analyzing video footage and user-defined criteria such as time and location.
Step 2: AcuSearch then pinpoints a specific object (e.g., a grey Audi SUV) across all scenes, providing a timeline of its appearances and related footage for further check.