| Title |
Multi-scale U-Net for Periodic Motion Detection |
| Authors |
유철환(Cheol-Hwan Yoo) ; 김호원(Ho-Won Kim) ; 한병옥(Byung-Ok Han) ; 장재윤(Jae-Yoon Jang) ; 유장희(Jang-Hee Yoo) |
| DOI |
https://doi.org/10.5573/ieie.2021.58.12.35 |
| Keywords |
Multi-scale U-Net; 3D CNN; Periodicity; Repetition |
| Abstract |
Recently, techniques for detecting repetitive and periodic segments in a video have been extensively studied in the field of computer vision. Conventional methods typically generate and utilize a self-similarity matrix as an intermediate representation for identifying repetitive segments in a video. However, these methods rely on a single-scale self-similarity matrix(SSM) and thus have a limitation that classification accuracy drops for videos including repetitive segments with various lengths and scales. To solve these problems, the encoder of the proposed network firstly generates self-similarity matrices, which incorporate information on various temporal scales by utilizing feature vectors extracted from multiple layers of the 3D CNN. By providing generated self-similarity matrices as input of a multi-scale U-Net through a multi-scale feature ensemble module, repetitive segments of various lengths in the video can be efficiently detected. Extensive experiments on the Countix and PERTUBE datasets demonstrate that the proposed network not only outperforms most hand-craft feature-based methods but also the latest deep learning-based methods. |