Action Recognition using Convolutional Neural Networks with Joint Supervision
Yupeng Li, Yuxiao Wang, Yongfeng Jiang, Liang Zhang · 2019
Mapping the depth video into an optimally representation in two-dimensional space are of vital importance for depth video based human action understanding. Meanwhile, such representation will lost some useful information inevitably, a feature learning approach not only separable but also discriminative are essential for action recognition task from such representation. This paper presents a new method for action recognition base on convolutional neural networks with joint supervision which shares the merits of both representation as mentioned above and convolutional neural networks. The advantages of our method come from (i) The whole procedure of our method is done automatically no matter the generation of representation or deeply feature learned; (ii) The deeply feature using the proposed deep architectures to learned has high discriminative capacity to improve the accuracy of action recognition effectively compared with handcrafted features. We conduct experiments on two challenging datasets: MSRDailyActivity3D and SYSU 3D HOI. Experimental results show that our method outperform previous methods based on hand-crafted features. Our method also achieves superior performance to the state-of-the-art on these datasets.