BTM: Boundary Trimming Module for Temporal Action Detection

Abstract

Temporal action detection (TAD) aims to recognize actions as well as their corresponding time spans from an input video. While techniques exist that accurately recognize actions from manually trimmed videos, current TAD solutions often struggle to identify the precise temporal boundaries of each action, which are required in many real applications. This paper addresses this problem with a novel Boundary Trimming Module (BTM), a post-processing method that adjusts the temporal boundaries of the detected actions from existing TAD solutions. Specifically, BTM operates based on the classification of frames in the input video, aiming to detect the action more accurately by adjusting the surrounding frames of the start and end frames of the original detection results. Experimental results on the THUMOS14 benchmark data set demonstrate that the BTM significantly improves the performance of several existing TAD methods. Meanwhile, we establish a new state of the art for temporal action detection through the combination of BTM and the previous best TAD solution.

Keywords

• action detection
• boundary trimming module
• video analytics