Vision Intelligence Management for Dynamic Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Electronic devices face limitations in simultaneously activating multiple recognition models due to limited computation resources, hindering the simultaneous performance of vision processing tasks like object recognition and scene analysis.
Innovation Solution
A vision intelligence management system that dynamically selects and activates models based on contextual information and resource optimization, allowing for parallel execution of multiple models using on-device and cloud processing, and supports both software and hardware acceleration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple recognition models are activated simultaneously for vision processing tasks, then recognition capabilities and analysis accuracy are improved, but computation resource consumption increases beyond device limits
Solution Approach 1:
The system dynamically adjusts which models are active based on runtime conditions. The execution manager monitors device state and contextual information to selectively activate models, transitioning between different model sets as conditions change, resolving the contradiction between maintaining high recognition accuracy and managing computation resource consumption
Solution Approach 2:
The system changes operational parameters by switching between different model configurations based on context. Depending on the visual data domain and device state, the system activates appropriate subsets of models, changing the active model parameters to match requirements while staying within resource constraints
2Adaptability or versatility
If all available recognition models are made active for comprehensive vision analysis, then vision processing capabilities are enhanced, but device performance and responsiveness deteriorate due to resource constraints
Solution Approach 1:
The model ecosystem is segmented into multiple domains (e.g., natural scenes, urban environments, medical imaging). Instead of activating all models simultaneously, the system segments the task by identifying the relevant domain and activating only the models appropriate for that domain, maintaining versatility while preserving device performance
Solution Approach 2:
The execution manager provides universal control over model activation across different domains. A single management system handles diverse vision processing tasks by selectively activating appropriate models, achieving multi-functionality without requiring all models to run simultaneously, thus maintaining device performance
3Use of energy by moving object
If recognition models are activated based on contextual information and dynamic selection, then resource utilization is optimized, but system complexity increases due to model management overhead
Solution Approach 1:
The execution manager acts as an intermediary between the model ecosystem and applications. It handles the complexity of contextual analysis and model selection internally, presenting a simplified interface to applications while optimizing resource utilization through intelligent model activation based on contextual information
Data Source
Figure 1
Figure 2
Figure 3
AI summary
One embodiment provides a method comprising classifying one or more objects present in an input comprising visual data by executing a first set of models associated with a domain on the input. Each model corresponds to an object category. Each model is trained to generate a visual classifier result relating to a corresponding object category in the input with an associated confidence value indicative of accuracy of the visual classifier result. The method further comprises aggregating a first set of visual classifier results based on confidence value associated with each visual classifier result of each model of the first set of models. At least one other model is selectable for execution on the input based on the aggregated first set of visual classifier results for additional classification of the objects. One or more visual classifier results are returned to an application running on an electronic device for display.