Language-Guided Visual Search for Fine-Grained Content Localization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video analysis and understanding approaches struggle to efficiently and accurately locate specific visual content due to limitations in integrating language models with visual search, particularly in navigating complex visual spaces and overlooking fine-grained details.

Innovation Solution

A computer-implemented method leveraging a language model with common sense linguistic reasoning capabilities to guide visual search by deducing identifiers for visual space portions, interacting with object detectors and VQA models, and employing logical step-by-step reasoning to enhance efficiency and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If automated visual search methods are used to handle increasing video content volume, then retrieval efficiency is improved, but accuracy in locating specific visual content deteriorates due to limitations in integrating language models with visual search

Engineering Contradiction:
Improveretrieval efficiencyVSAvoidaccuracy in locating visual content
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent introduces a language model as an intermediary between the visual search system and the target visual content. The language model receives natural language queries and translates them into visual search parameters, acting as a mediator that bridges linguistic understanding and visual content retrieval. This intermediary enables more accurate localization of specific visual content while maintaining efficient automated search capabilities.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If language models are integrated with visual search to improve accuracy, then success rate increases, but system complexity worsens due to limitations in navigating complex visual spaces

Engineering Contradiction:
Improvesuccess rate of visual searchVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the visual search process into distinct components: a language model processing module that handles natural language queries, a visual space navigation module that manages the complex visual environment, and a retrieval module that locates target content. By dividing the system into specialized segments, each component can be optimized independently, reducing overall system complexity while maintaining high success rates.

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If visual search inspects more portions of visual space to improve accuracy, then localization precision is improved, but time consumption worsens

Engineering Contradiction:
Improvelocalization precisionVSAvoidtime consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent employs preliminary action by using the language model to pre-process and interpret natural language queries before initiating the visual search. The language model generates preliminary search parameters and constraints that guide the visual space inspection, allowing the system to focus on relevant portions of the visual content. This preliminary processing reduces the overall search space that needs to be inspected, achieving high localization precision with reduced time consumption.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4617912A1Computer-implemented methods and system for visual search
Publication Date: 2025.09.17 TOYOTA JIDOSHA KK
  • EP4617912A1 patent drawingFigure 1
  • EP4617912A1 patent drawingFigure 2
  • EP4617912A1 patent drawingFigure 3

AI summary

A computer-implemented method for visual search is provided. The method includes: transmitting to a language model a textual description describing a task to be completed by the language model through searching a visual space; instructing the language model to complete the task by selectively inspecting at least one portion of the visual space, wherein the inspection includes deducing, based on the textual description, an identifier identifying one of the at least one portion within the visual space, requesting a visual description of the portion from the computer by returning the identifier, and determining whether the task is completed based on the visual description; receiving the identifier from the language model; generating the visual description; and transmitting the visual description to the language model, so that the language model outputs a final response indicating an execution result of the task after determining whether the task is completed.