Video Clip Extraction via Siamese Scoring Model

The patent applies preliminary action by pre-labeling video clips with target attributes (such as highlight, ordinary, or transition clips) before training the model. This preprocessing step creates clean, structured data pairs that guide the model training process, allowing the model to learn effective clip selection without requiring complex network structures or extensive training time. The labeled data pairs are prepared in advance to reflect user preferences, enabling the model to quickly converge to accurate clip extraction performance.

The patent extracts and removes the noisy data from the training process by using pre-labeled clips with target attributes. Instead of attempting to handle noisy data through complex network structures, the solution extracts clean, labeled data pairs that directly represent the desired clip selection criteria. This extraction of clean data eliminates the need for complex noise-handling mechanisms while maintaining high accuracy in clip selection.

The patent applies local quality by differentiating between different types of clips and assigning specific target attributes to each type. Instead of treating all video data uniformly, the solution labels clips locally with specific attributes (highlight, ordinary, transition) based on their characteristics. This localized quality assignment ensures that only relevant, high-quality labeled data is used for training specific clip types, improving both model accuracy and data quality by avoiding the dilution effect of noisy, unlabeled data.