As an important part of video understanding, temporal action detection (TAD) has wide application scenarios. It aims to simultaneously predict the boundary position and class label of every action instance in an untrimmed video. Most of the existing temporal action detection methods adopt a stacked convolutional block strategy to model long temporal structures. However, most of the information between adjacent frames is redundant, and distant information is weakened after multiple convolution operations. In addition, the durations of action instances vary widely, making it difficult for single-scale modeling to fit complex video structures. To address this issue, we propose a non-local temporal difference network (NTD), including a chunk convolution (CC) module, a multiple temporal coordination (MTC) module, and a temporal difference (TD) module. The TD module adaptively enhances the motion information and boundary features with temporal attention weights. The CC module evenly divides the input sequence into N chunks, using multiple independent convolution blocks to simultaneously extract features from neighboring chunks. Therefore, it realizes the information delivered from distant frames while avoiding trapping into the local convolution. The MTC module designs a cascade residual architecture, which realizes the multiscale temporal feature aggregation without introducing additional parameters. The NTD achieves a state-of-the-art performance on two large-scale datasets, 36.2% mAP@avg and 71.6% mAP@0.5 on ActivityNet-v1.3 and THUMOS-14, respectively.
Keywords: computer vision; convolutional neural networks; deep learning; temporal action detection; video understanding.