Video Rule Generation Using Pre-Trained Model Attribute Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and labor-intensive process of preparing and labeling training data for deep learning models, particularly in the context of video scene recognition, where thousands or tens of thousands of data points are required, making it difficult to implement effective video search systems.
Innovation Solution
An information processing apparatus that uses pre-trained models to identify attribute labels in sample videos, clusters these labels based on position and time, and generates rules from the most frequent clusters, reducing the need for extensive training data and labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning model learning is performed for video scene recognition, then recognition accuracy is improved, but training data cost increases significantly
Solution Approach 1:
The patent applies preliminary action by using pre-trained models to extract attribute labels from sample videos before generating recognition rules. Instead of requiring extensive training data for deep learning, the system first pre-processes a small set of sample videos through pre-trained models to obtain attribute information, which is then used to construct recognition rules. This preliminary extraction of useful information from minimal data resolves the contradiction between achieving high recognition accuracy and reducing training data quantity.
2Measurement precision
If deep learning model learning is performed for video scene recognition, then recognition accuracy is improved, but labeling time increases
Solution Approach 1:
The patent applies self-service by using pre-trained models to automatically extract attribute labels from sample videos without requiring manual labeling. The system feeds sample videos into pre-trained models, which automatically generate attribute information such as object presence, actions, and scene characteristics. This automated attribute extraction eliminates the need for time-consuming manual labeling while maintaining high recognition accuracy through the generated rules.
3Measurement precision
If attribute labels are clustered based on position and time, then rule generation accuracy is improved, but processing complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the attribute label processing into distinct stages: first extracting attributes from sample videos, then clustering labels based on position and time characteristics, and finally generating recognition rules from the clustered results. This segmentation of the processing pipeline allows the system to handle complexity in a structured manner, improving rule generation accuracy through systematic clustering while managing processing complexity through modular organization of operations.
Data Source
AI summary
An information processing apparatus includes: a memory; and a processor coupled to the memory and the processor configured to: acquire a plurality of sample videos; identify a position and time at which an attribute appears in each of the plurality of sample videos, the attribute being output by each of one or more pre-trained models to which each of the plurality of sample videos is input; cluster attribute labels based on the position and time of the attribute for each of the plurality of sample videos; and generate a rule by combining attribute labels included in a cluster having a highest frequency of appearance among cluster groups obtained for all of the plurality of sample videos.