Supporting One-Time Point Annotations for Gesture Recognition.

Area of Science:

• Human-Computer Interaction
• Machine Learning
• Signal Processing

Background:

• Gesture recognition relies on accurately annotated training data, which is time-consuming to create.
• Conventional annotation requires precise start/end times and labels for each gesture.
• Reducing annotation burden is crucial for developing practical gesture recognition systems.

Purpose of the Study:

• To propose and evaluate a novel, time-efficient annotation technique for gesture recognition.
• To introduce an algorithm for automatically refining imprecise annotations into accurate temporal boundaries.
• To assess the performance of models trained with the new annotation method compared to traditional approaches.

Main Methods:

• A one-time point annotation strategy was developed, simplifying the labeling process.
• The BoundarySearch algorithm was proposed to automatically identify gesture boundaries from single-point annotations.
• Gesture recognition models were trained using the corrected annotations and evaluated on multiple datasets.

Main Results:

• Training models with corrected one-time point annotations achieved performance comparable to fully supervised methods (within 5% F1-score difference on average).
• The BoundarySearch algorithm demonstrated effectiveness across various wearable gesture recognition applications and sensor modalities.
• The algorithm also showed strong performance on the ChaLearn 2014 multi-modal gesture recognition challenge.

Conclusions:

• The proposed one-time point annotation technique and BoundarySearch algorithm significantly reduce annotation effort without substantial performance loss.
• This approach offers a practical solution for creating large-scale gesture recognition datasets.
• The method is robust and applicable to diverse gesture recognition tasks and data types.