Abstract
At present, algorithms for human motion capture and 3D human reconstruction are not perfect in some experiment, such as PIFu-HD. In addition, there are some reconstruction errors in practical application. We believe that there is background interference. In this paper, we provide the same action in different backgrounds for PIFu-HD model. We endow a single RSU block of U2 Net with multi-receptive field mechanism, and use it for dataset of salient object detection (SOD), achieving a significant improvement in accuracy. Then we do the comparative experiment which provides an opportunity of analysing the possible causes of errors in existing algorithms. Thus we propose a new method, a background removal based PIFu-HD. In the end, we use well-constructed dance images and videos for relevant modelling and comparison. We also present the well-constructed dataset and a formulated standard for Cha-Cha, which is of great importance to model-training.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have