Action Recognition Model Parameter Adjustment via Synthetic Marker Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine learning methods for action recognition in videos face challenges in collecting enough marker samples, leading to overfitting and reduced accuracy due to the need for extensive training data and resource consumption.
Innovation Solution
A model-based data processing method that extracts a second training sample set and query video from a first training sample set, processes them through an embedded layer network and time sequence relation network to obtain frame feature sequences and descriptors, which are then used to adjust the action recognition model parameters, enhancing generalization and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning methods are used for action recognition, then the model can be trained with available data, but the model suffers from overfitting due to insufficient marker samples
Solution Approach 1:
The patent applies preliminary action by pre-processing video data to automatically generate synthetic marker samples before actual model training. The system extracts action patterns from unlabeled videos and creates artificial training data with simulated marker points, enabling the model to learn from expanded datasets without requiring manual collection of sufficient real marker samples.
Solution Approach 2:
The patent uses copying by creating synthetic copies of action patterns through video processing. The system generates multiple virtual marker sample instances by detecting and replicating action trajectories from source videos, effectively multiplying the available training data without physical collection of additional real-world samples.
2Measurement precision
If extensive training data is collected for traditional machine learning, then the model accuracy may improve, but the resource consumption and time required for data collection and processing increase significantly
Solution Approach 1:
The patent replaces the mechanical process of manual data collection and labeling with an automated computer vision system. The system uses video processing algorithms and neural networks to automatically detect actions, generate marker points, and create training datasets, eliminating the time-consuming manual work while maintaining or improving data quality and model accuracy.
Solution Approach 2:
The system applies self-service by enabling automatic self-generation of training data from video inputs. The model autonomously processes video content, identifies action patterns, and produces synthetic training samples without external human intervention, making the data preparation process self-sufficient and highly efficient.
3Reliability
If more marker samples are collected to avoid overfitting, then the model generalization improves, but the complexity of data collection and processing increases
Solution Approach 1:
The patent applies universality by designing a multi-functional video processing system that simultaneously performs action detection, marker point generation, and training data creation. This integrated approach consolidates multiple complex functions into a unified framework, reducing overall system complexity while generating sufficient diverse training samples for improved model generalization.
Data Source
AI summary
A model-based data processing method and apparatus that can extract a first training sample set to obtain a second training sample set and a query video, the first training sample set including different types of video samples; process the second training sample set through an embedded layer network in an action recognition model to obtain a first frame feature sequence; process the query video through the embedded layer network to obtain a second frame feature sequence; process the first frame feature sequence through a time sequence relation network in the action recognition model to obtain a first time sequence relation descriptor; process the second frame feature sequence through the time sequence relation network to obtain a second time sequence relation descriptor; and adjust a model parameter of the action recognition model according to the first time sequence relation descriptor and the second time sequence relation descriptor.