Desktop Action Recognition From First-Person Point-of-View

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Action recognition from first-person (egocentric) video is crucial for understanding daily activities, especially hand-object interactions.
• Existing research lacks a dedicated benchmark dataset for desktop action recognition tasks.

Purpose of the Study:

• To introduce a novel benchmark dataset for egocentric desktop action recognition.
• To propose a feature representation for characterizing desktop actions based on hand movements.
• To establish baseline performance for desktop action recognition.

Main Methods:

• Recorded egocentric videos of six participants performing common desktop actions using a head-mounted camera.
• Developed frame-level action annotations and pixel-level hand masks for the dataset.
• Proposed a feature representation leveraging spatial and temporal hand information.

Main Results:

• Released a new benchmark dataset for egocentric desktop action recognition.
• Demonstrated the dataset's statistical information and baseline performance.
• The proposed feature representation achieved promising results for recognizing five distinct action classes.

Conclusions:

• The developed dataset and proposed feature representation offer a valuable resource for advancing egocentric desktop action recognition research.
• The findings highlight the potential of analyzing hand-object interactions in first-person videos for improved action recognition.