StrokeRehab: A Benchmark Dataset for Sub-second Action Identification.

Area of Science:

• Machine Learning
• Computer Vision
• Biomedical Engineering

Background:

• Current action identification models struggle with elemental, short-duration motions critical for applications like stroke rehabilitation.
• Existing datasets often lack the high temporal resolution needed for fine-grained action recognition.

Purpose of the Study:

• To address the limitations in recognizing elemental actions at high temporal resolution.
• To introduce a large-scale, multimodal dataset (StrokeRehab) for action recognition benchmarking.
• To develop a novel approach for accurate high-resolution action identification.

Main Methods:

• Developed the StrokeRehab dataset with video and inertial measurement unit (IMU) sensor data from healthy and stroke-impaired individuals performing daily activities.
• Proposed a novel sequence-to-sequence model inspired by speech recognition for direct action sequence prediction.
• Evaluated the proposed model on StrokeRehab and standard benchmark datasets (50Salads, Breakfast, Jigsaws).

Main Results:

• The StrokeRehab dataset enables studying distribution shift in action recognition due to healthy vs. impaired subject data.
• State-of-the-art models showed noisy predictions on StrokeRehab, highlighting the need for new approaches.
• The proposed sequence-to-sequence model significantly outperformed existing methods on StrokeRehab and other benchmarks.

Conclusions:

• The StrokeRehab dataset is a valuable resource for advancing high-resolution action recognition research.
• The novel sequence-to-sequence approach effectively addresses the challenge of identifying elemental actions with high temporal precision.
• This work has significant implications for smart health applications, particularly in personalized rehabilitation and activity monitoring.