Mastering Computer Vision with TensorFlow 2.x

Combining video-based actions with pose estimation

Action recognition can be two-dimensional and three-dimensional. The two-dimensional action recognition methods use the joint information of the body, which is represented by key points. These key points are represented in a vector called a feature map. On the other hand, the three-dimensional action recognition methods need not only the feature map but the whole body's skeleton data. This data can be obtained using depth sensors such as Microsoft Kinect or Intel RealSense. In 2018, Diogo C. Luvizon, David Picard, and Hedi Tabia introduced their paper titled 2D/3D Pose Estimation and Action Recognition using Multitask Deep Learning. The details of this paper can be found at https://arxiv.org/abs/1802.09232 ...