Probabilistic Temporal Modeling for Unintentional Action Localization.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Localizing unintentional actions in videos is challenging due to unreliable annotations stemming from subjective appraisals and ambiguity.
• Existing methods struggle with stable training due to the inherent difficulty in obtaining consistent labels for unintentional actions.

Purpose of the Study:

• To develop a novel probabilistic framework for unintentional action localization that effectively models annotation uncertainty.
• To address the limitations of subjective and ambiguous annotations in video-based action recognition.

Main Methods:

• Proposed a probabilistic framework with two components: Temporal Label Aggregation (TLA) and Dense Probabilistic Localization (DPL).
• TLA formulates annotated failure moments as temporal label distributions and aggregates them online for dense probabilistic supervision.
• DPL jointly trains probabilistic dense classification, temporal detection, and regression heads with varying supervision granularities for collaborative learning.

Main Results:

• The proposed framework demonstrated significant improvements over baseline and state-of-the-art methods on the OOPS dataset.
• Effective modeling of annotation uncertainty led to more robust unintentional action localization.

Conclusions:

• The probabilistic framework successfully addresses the challenge of unreliable annotations in unintentional action localization.
• This work advances the capability of machines to understand and localize human actions, particularly unintentional ones, in video data.