Combination of Deep Learner Network and Transformer for 3D Human Pose Estimation

Deep neural networks (DNNs) have attained the maximum performance today not just for human pose estimation but also for other machine vision applications (e.g., semantic segmentation, object detection, image classification). Besides, the Transformer shows its good performance for extracting the information in temporal information for video challenges. As a result, the combination of deep learner and transformer gains a better performance than only the utility one, especially for 3D human pose estimation. At the start point, input the 2D key point into the deep learner layer and transformer and then use the additional function to combine the extracted information. Finally, the network collects more data in terms of using the fully connected layer to generate the 3D human pose which makes the result increased precision efficiency. Our research would also reveal the relationship between the use of the deep learner and transformer. When compared to the baseline-DNNs, the suggested architecture outperforms the baseline-DNNs average error under Protocol 1 and Protocol 2 in the Human3.6M dataset, which is now available as a popular dataset for 3D human pose estimation.