Video Motion Labeling With Combined Labels at Temporal Boundaries

The patent divides the video into multiple sections based on motion type boundaries and selects representative frames from each section for annotation. This segmentation approach allows annotators to focus on key temporal points rather than annotating every frame, significantly reducing work cost while maintaining training data accuracy through the combined label generation process.

The patent introduces a combined label generation process that acts as an intermediary between the sparse timestamp annotations and the full frame labels. By combining labels from representative frames with predicted labels from the machine learning model, the system creates reliable training data without requiring manual annotation of all frames, thus reducing work cost while maintaining accuracy.

The patent implements a feedback mechanism where the machine learning model predicts labels for frames between representative frames, and these predictions are combined with the timestamp annotations to generate combined labels. This feedback loop allows the system to leverage both the reduced annotation effort and the model's predictive capabilities to maintain high label reliability despite using sparse timestamp annotations.

The patent merges timestamp annotations with machine learning model predictions to generate combined labels. This combination approach integrates the reliability of manually annotated representative frames with the computational efficiency and predictive accuracy of the machine learning model, resulting in high-quality training data that maintains reliability while reducing work cost.

The patent performs preliminary action by selecting representative frames that capture the essence of each motion section before generating the combined labels. These representative frames serve as anchors that guide the machine learning model in predicting labels for intermediate frames, ensuring that even frames near temporal boundaries receive accurate labeling through the combined label generation process.

The patent changes the parameter of label generation from binary (annotated or not) to a combined approach that integrates multiple sources of information. By transforming the labeling process into a parameter-rich approach that combines timestamp annotations with model predictions, the system maintains high estimation accuracy at temporal boundaries while preserving data preparation efficiency.