.png){kind=icon}
.png){kind=icon}
Return
Bottom-up improved multistage temporal convolutional network for action segmentation
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03382-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Action segmentation involves locating and classifying human action segments in an untrimmed video, which is very important for understanding human activities. Segmenting actions in the video is a very challenging task due to the problem of ambiguous frames. Previous studies on this topic usually required additional inputs or constructed highly complicated network structures to achieve good performance. However, these additional inputs are not easy to obtain, and complicated network structures increase the costs of computation and storage. Hence, to mitigate these problems, we propose a bottom-up improved multistage temporal convolutional network (BUIMS-TCN) for action segmentation. Specifically, we first propose a smoothed dilated 1D convolution to learn the inherent local temporal dependencies. Second, we design an adaptive temporal fusion module (ATFM), which is a simple yet effective multiscale temporal-context information fusion module, to obtain better semantic feature representations. Finally, we introduce a new loss function to solve the imbalance between easy and hard samples. To the best of our knowledge, this is the first time that the above improvements have been incorporated into the action segmentation task. Extensive experiments verify that our model significantly outperforms the state-of-the-art baselines on three challenging benchmark datasets: Georgia Tech Egocentric Activities (GTEA), 50Salads, and the Breakfast dataset.
Keywords:
Action segmentation
Smoothed dilated 1D convolution
Adaptive temporal fusion module
Temporal convolutional network
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W