Temporal Action Detection Based on Temporal Convolutional Network with Coarse-Grained Clips Construction

Abstract

Temporal action detection has gradually become a hot field with the development of video. This task focuses on untrimmed video applications, which is close to practical applications. To improve the accuracy of temporal action detection, this paper proposes a novel framework based on temporal convolutional network with coarse-grained clips construction and our framework mainly includes three parts as follows: feature extraction, coarse-grained clips construction and improved boundary regression model. In particular, to make our model more stable without reducing the sensitivity to boundary, this paper proposes a novel coarse-grained clips construction to gain coarse-grained clips. Furthermore, we refine action boundary with improved boundary regression model based on temporal convolutional network and design three sub-module structures to preserve temporal information. The results on THUMOS14 illustrate that our framework could provide high quality of action propose to improve the detection accuracy.