Hybrid ML Architecture for Custom Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video monitoring technologies face challenges in efficiently processing large volumes of media content and managing custom object classification and tracking, especially in scenarios where computing resources are limited and custom objects need to be identified without extensive initial training data.
Innovation Solution
A hybrid machine learning architecture using teacher and student models to create a custom model for identifying user-defined objects, where the student model selects training candidates based on search configuration parameters, and the custom model is deployed on edge devices for initial analysis, with advanced models in the cloud performing detailed analysis only on selected, interesting content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If manual selection and labeling of training samples is used for custom object classification, then custom objects can be identified, but the process becomes cumbersome and time-consuming
Solution Approach 1:
The system performs self-service by automatically selecting and labeling training samples using the student model and teacher models. The student model identifies candidate samples, and the teacher models automatically generate labels without human intervention, enabling the system to adapt to custom objects autonomously
Solution Approach 2:
The system performs preliminary actions by pre-selecting training candidates using the student model before the actual training process. This preliminary selection of relevant samples accelerates the overall training process and reduces the time required for custom object classification setup
2Measurement precision
If all media content is processed by advanced machine learning models, then accurate analysis is achieved, but computational resources are excessively consumed
Solution Approach 1:
The processing system is segmented into two levels: edge devices perform initial analysis using lightweight custom models, and only selected interesting content is forwarded to cloud-based advanced models for detailed analysis. This segmentation distributes computational load and reduces overall resource consumption while maintaining accuracy
Solution Approach 2:
Instead of applying advanced models to all media content, the system applies them only partially to selected content that meets certain criteria. This partial action approach maintains high accuracy for relevant content while significantly reducing computational resource consumption
3Measurement precision
If extensive training data is collected for custom object recognition, then model accuracy improves, but data exposure and security risks increase
Solution Approach 1:
The training process is segmented into local and cloud components. Local edge devices perform initial training using collected samples, and only selected interesting content and model updates are transmitted to the cloud. This segmentation minimizes data exposure while maintaining recognition accuracy
Solution Approach 2:
The system implements local quality by enabling custom object training to occur locally at edge devices using collected samples. This local processing capability allows the system to achieve accurate custom object recognition without requiring extensive centralised data collection and storage, thereby reducing data exposure risks
Data Source
AI summary
Systems and methods for visual content processing. A method includes applying teacher models to training candidates in order to output instances of a custom object label. The training candidates are selected using a student model based on search configuration parameters. A first set of media content is generated by labeling the training candidates based on the instances of the custom object label output by the teacher models. A custom model is created using the teacher models. The custom model is a machine learning model trained using the first set of media content. A subset of a second set of media content is obtained. The subset of the second set of media content is selected based on outputs of the custom model as applied to the second set of media content. An advanced machine learning model is applied to the obtained subset of the second set of media content.