Abstract
The task of naturalistic driving action localization carries significant safety implications, as it involves detecting and identifying possible distracting driving behaviors in untrimmed videos. Previous studies have demonstrated that action localization using a local snippet followed by probability-based post-processing, without any training cost or redundant structure, can outperform existing learning-based paradigms. However, the action probability is computed at the snippet-level, the input information near the boundaries is attenuated, and the snippet size is limited, which does not support the generation of more precise action boundaries. To tackle these challenges, we introduce an action probability calibration module that expands snippet-level action probability to the frame-level, based on a preset snippet position reliability, without incurring additional costs for probability prediction. The frame-level action probability and reliability enable the use of various snippet sizes and equal treatment for information of different temporal points. Additionally, based on the calibrated probability, we further design a category-customized filtering mechanism to eliminate the redundant action candidates. Our method ranks 2nd on the public leaderboard, and the code is available at https://github.com/RongchangLi/AICity2023_DrivingAction.