Deep Neural Network for Sequence Labeling of Motion Data

Description:

In this work, we develop a new skeleton-based action recognition framework based on convolutional neural networks. In particular, we devise new type of convolutional filters which are more suitable for temporal input data and can extract semantic parts of the action motion. These filters lead to learning temporal prototypes based on which the action categories can be effectively classified/segmented. In addition, the prediction outcome of the network is interpretable based on those extracted prototypes.

In addition, we designed our action-recognition framework flexible to input length (temporal) and developed an incremental extension module which increase the network depth according to the complexity of the input data.

The implementations on well-known action recognition datasets shows the high performance of our method compared to the state-of-the-arts and illustrates its interpretable trained filters .

The detail description of the framework will be publicly available after the reviewing process the submitted paper is concluded. However, please contact me regarding the private distribution of the work.