EventAnchor: Reducing Human Interactions in Event Annotation of Racket Sports Videos

Title of the Paper

EventAnchor: Reducing Human Interactions in Event Annotation of Racket Sports Videos

Paper Information

• Subject Area: Sports video analysis and AI-assisted annotation
• Keywords: sports video, visual analysis, human-computer interaction, data annotation, event detection, table tennis

Research Background and Problem Statement

• Problems and Challenges:

  • Manually extracting key information from lengthy match videos is highly challenging.
  • Existing computer vision-based data collection systems perform poorly on low-quality videos (e.g., broadcast videos), making it difficult to accurately track objects.
  • These systems often focus on low-level object recognition (e.g., human actions) but are weak in extracting high-level event information (e.g., action outcomes).
  • Current annotation systems lack scalability and are inefficient when handling large-scale annotation tasks.
  • For fast-paced sports like table tennis, existing data annotation methods are not effectively applicable.

• Significance of the Research:

  • Table tennis and other racket sports are globally popular, with a high demand for data analysis, particularly for match tactics and player performance evaluation.

• Motivation and Related Work:

  • Experts from different fields require analysis of table tennis match videos, including ball and player positions, action types, and tactical styles. Existing systems fail to meet these complex analytical needs.
  • The paper reviews current interactive video annotation systems, model-based annotation methods, and their design challenges.

Proposed Solution

• Proposed Solution:

  • The authors propose the EventAnchor framework, which uses computer vision algorithms to identify significant events as anchors to help users locate, analyze, and annotate match videos.
  • They designed a table tennis annotation system based on EventAnchor, named EventAnchor for Table Tennis (ETT).

• Innovations:

  • Dividing annotation tasks into three levels:
    • Object Level: Identifying key objects in the video (ball, players, table, etc.) using computer vision.
    • Event Level: Capturing significant events (e.g., racket-ball contact, ball bouncing on the table) through object motion information and specific algorithms.
    • Context Level: Allowing users to add contextual information interactively (e.g., tactical types).
  • Enhancing the timeline tool for quick video event browsing, with the use of a "calibration box" and "annotation box" to assist users in efficient labeling.

• Technical Implementation:

  • Applied the FOTS model for optical character recognition (OCR).
  • Used TrackNet to extract ball trajectories and OpenPose to estimate player poses.
  • Employed ResNet-50 and support vector machines to classify video scenes.
  • Provided interactive tools for time calibration and event annotation.

Research Outcomes

• Specific Outcomes:

  • The framework's effectiveness was validated through two experiments:
    1. Event-Level Task: Annotating the time and location of ball-table contact in table tennis videos.
    2. Context-Level Task: Labeling the "serve and attack" tactical type used by players during matches.

• Experimental Results:

  • Compared to traditional systems, ETT significantly improved annotation efficiency:
    • Event-Level Task: Average task completion time reduced by approximately 14%, with a 50% reduction in time errors.
    • Context-Level Task: Completion time for complex tactical annotations was reduced by over 50%.
  • Users provided positive feedback on ETT's design and functionality, noting that anchor hints and interactive tools greatly reduced cognitive and interaction burdens.

• Comparison with Existing Solutions:

  • ETT outperformed baseline systems in annotation efficiency.
  • For complex context-level tasks, ETT demonstrated significant time advantages, though annotation accuracy was comparable to baseline systems.

• Limitations and Future Directions:

  • Limitations:
    • Limited flexibility in event definitions, currently relying primarily on ball contact with the table or racket.
    • Insufficient utilization of audio information.
    • Users may lack sufficient validation of results provided by computer vision.
  • Future Work:
    • Develop functionality for users to define new event types.
    • Enhance audio information recognition to complement visual data.
    • Improve interaction design to increase user engagement and result validation accuracy.