Computer Vision Scene Estimation Using Dual Machine Learning Models

The patent divides the estimation task into multiple segments by creating separate machine learning models for different scene type lengths. Instead of using a single uniform model, the system segments the problem into: (1) a first machine learning model for short scene types (e.g., goal scenes), (2) a second machine learning model for long scene types (e.g., corner kick scenes), and (3) a determination unit that selects the appropriate model based on scene characteristics. This segmentation resolves the contradiction by maintaining training simplicity through dedicated models while improving accuracy through length-specific training data.

The patent applies local quality by training different machine learning models with different characteristics suited to specific scene type lengths. The first machine learning model is trained on short period data with specific feature extraction parameters, while the second machine learning model is trained on long period data with different parameters. This allows each model to have optimized local quality for its specific scene type, resolving the contradiction between uniform training simplicity and accurate local estimation.

The patent segments the estimation system into multiple specialized machine learning models, each handling specific scene type lengths. The determination unit then segments the classification decision by selecting the appropriate model based on detected scene characteristics. This segmentation approach improves accuracy by dedicating specific models to specific tasks while managing complexity through a clear hierarchical structure.

The patent introduces dynamics by making the model selection process adaptive rather than static. The determination unit dynamically selects which machine learning model to use based on the detected scene characteristics and length. This dynamic approach allows the system to optimize accuracy for each specific scene type while managing overall system complexity through intelligent model routing rather than requiring all models to handle all cases.

The patent segments the training process into separate, focused tasks for different scene type lengths. Each machine learning model is trained independently on specialized data, which allows for more efficient training compared to a single model trying to handle all scene types. The determination unit then segments the inference process by selecting the appropriate pre-trained model, reducing real-time processing time.

The patent applies preliminary action by pre-training multiple machine learning models with different length specifications before actual estimation is needed. This allows the system to prepare specialized models in advance, so that during actual scene type estimation, the system only needs to select the appropriate pre-trained model rather than training new models, significantly reducing inference time while maintaining high accuracy.