Multimodal Scene Recognition with Structured Memory Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing scene recognition approaches for autonomous systems are complex, time-consuming, and energy-demanding, requiring multiple trained models for variable tasks, and struggle with complex environments not covered by their training, leading to poor performance and short-term analysis of unstructured information.

Innovation Solution

A system utilizing a Large Multimodal Model (LLM) with Visual Language Model (VLM) and Large Language Model (LLM) functionalities, combined with a structured memory to generate an enhanced scene description by comparing first and second scene descriptions, incorporating past observations and current data for improved contextual understanding.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple trained models are used for variable tasks, then task coverage is improved, but system complexity and computational cost increase

Engineering Contradiction:
Improvetask coverageVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent employs a single Large Multimodal Model that performs multiple functions including scene description generation, object recognition, and contextual understanding, replacing the need for multiple specialized trained models while maintaining versatility across different tasks

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If neural networks are trained for specific tasks, then task performance is improved, but energy consumption and training time increase

Engineering Contradiction:
Improvetask performanceVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The Large Multimodal Model is pre-trained on diverse datasets to acquire general scene understanding capabilities beforehand, allowing it to handle specific tasks without requiring extensive task-specific training, thereby reducing energy consumption and training time for each application

Inventive Principle:
Principle #10Preliminary action

3Speed

If feed forward recognition pipeline is used, then processing speed is improved, but contextual understanding and long-term analysis capability deteriorate

Engineering Contradiction:
Improveprocessing speedVSAvoidcontextual understanding
Core Design Contradiction:
SpeedVSLoss of information

Solution Approach 1:

The system incorporates feedback mechanisms where the Large Multimodal Model continuously refines scene descriptions by comparing generated descriptions with actual observations, enabling long-term analysis and contextual understanding while maintaining processing efficiency through iterative improvement rather than purely feed-forward processing

Inventive Principle:
Principle #23Feedback

4Ease of operation

If short-term analysis with unstructured information is performed, then processing simplicity is maintained, but analysis depth and adaptability deteriorate

Engineering Contradiction:
Improveprocessing simplicityVSAvoidanalysis depth
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system transforms unstructured visual information into structured scene descriptions with defined parameters and contexts, enabling deeper analysis and better adaptability while maintaining processing simplicity through standardized information representation formats

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4645266A1System for for generating information on a real world scene
Publication Date: 2025.11.05 HONDA MOTOR CO LTD
  • EP4645266A1 patent drawingFigure 1
  • EP4645266A1 patent drawingFigure 2
  • EP4645266A1 patent drawingFigure 3

AI summary

The invention is related to a system (1) for analyzing an image of a real world scene, the system (1) comprising image providing means (13) for obtaining the image of the real world scene, a large multimodal model LMM module (11) providing a large language model LLM functionality and a visual language model VLM functionality, wherein the LMM module (11) is configured to analyze the image using its VLM functionality for generating a first scene description of the received image, a structured memory (12) for storing real world information, wherein the structured memory (12) is connected to the LMM module (11) and configured to generate a second scene description of the received image based on the stored real world information and to provide the description to the LMM module (11), wherein the LMM module (11) is configured to identify differences between the first scene description and the second scene description and to generate an enhanced scene description based on the identified mismatches.