Temporal stochastic linear encoding networks
作者：
Highlights：
• The proposed TSLE network can implement an easy yet robust manipulation of long range temporal clues.
• We propose an arbitrarily directional motion boundary (ADMB) units embodied inside of the CNN, which can save the training time and hard disk space.
• We train the proposed TSLE networks on the large-scale Kinetics dataset, and with the obtained model we train the models on the UCF101 and HMDB51.
• The proposed TSLE networks map a highly-dimensional video to a compact spatio-temporal representation.
摘要
•The proposed TSLE network can implement an easy yet robust manipulation of long range temporal clues.•We propose an arbitrarily directional motion boundary (ADMB) units embodied inside of the CNN, which can save the training time and hard disk space.•We train the proposed TSLE networks on the large-scale Kinetics dataset, and with the obtained model we train the models on the UCF101 and HMDB51.•The proposed TSLE networks map a highly-dimensional video to a compact spatio-temporal representation.
论文关键词：Convolutional neural network,Action recognition,Long-range temporal dynamics,Motion boundary
论文评审过程：Received 16 March 2018, Revised 31 July 2018, Accepted 4 September 2018, Available online 6 September 2018, Version of Record 14 September 2018.
论文官网地址：https://doi.org/10.1016/j.image.2018.09.001
