Lu Dong (董璐)Ph.D. Student
Department of Computer Science and Engineering |
 |
A smooth 3D Human Pose Estimation pipeline for video in the wild
Abstract
The majority of current models have been trained on limited close set data, which often results in subpar performance when applied to real-world video data. One of the major challenges in this context is the issue of self-occlusion, which refers to instances where parts of a subject's body obstruct or cover other parts. This results in a less smooth and precise performance. To tackle this issue, my approach involved a transformer-based pipeline that incorporates both self communication and cross communication mechanisms. Applying this technique in consecutive frames has resulted in a marked improvement in smooth and has been verified through extensive experimentation. The results of these experiments demonstrate that the proposed pipeline outperforms state-of-the-art performance on the Human3.6 dataset, as well as exhibiting the best performance on our in-house self-occlusion dataset.
|
Last Updated on Feb, 2023
|
