Shandong Provincial Key Laboratory of Network Based Intelligent Computing, School of Information Science and Engineering, University of Jinan, Jinan, China
Jianqin Yin
Shandong Provincial Key Laboratory of Network Based Intelligent Computing, School of Information Science and Engineering, University of Jinan, Jinan, China
Laboratory for Human Machine Control, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
Huaping Liu
Department of Computer Science and Technology, Tsinghua University, Beijing, China
Qi Dang
Department of Computer Science and Technology, Tsinghua University, Beijing, China
Zhiming Li
Shandong Provincial Key Laboratory of Network Based Intelligent Computing, School of Information Science and Engineering, University of Jinan, Jinan, China
Yilong Yin
School of Software Engineering, Shandong University, Jinan, China
This research proposes a method for optimizing extracted candidate proposals based on the action temporal semantic continuity rule to accurately detect the category and start and end time in the temporal action detection of long untrimmed videos. First, sliding windows of the same scale and different scales are integrated according to the rule of action temporal semantic continuity. Subsequently, we reacquire the classification confidence score and relocate the integration results. Finally, inaccurate detections are eliminated by non-maximum value suppression. In contrast to the specified scale of the detection results obtained by sliding window, this method can produce the action temporal segments of any length and suppress the redundant detection. Therefore, the detection results are more consistent with the expectation of an individual. Experimental results show that the mean average precision increases from 19.0% to 20.6% when the intersection-over-union threshold is set to 0.5 on THUMOS 2014 data set.
