3D Pose Estimation With AI For Heavily Occluded Images
Overview of the paper “CenterHMR: a Bottom-up Single-shot Method for Multi-person 3D Mesh Recovery from a Single Image” by Y Sun et al.
In recent years we have seen a lot of pose estimation techniques that work near flawlessly even when trying to infer 3D pose information from just 2D images. And today we have yet another paper that improves upon this technique ever-so-slightly, making it more robust and closer to being useful in real-life applications. The paper I am talking about is titled “CenterHMR: a Bottom-up Single-shot Method for Multi-person 3D Mesh Recovery from a Single Image”.
This novel bottom-up approach improves detection of humans even in cases with heavy occlusion where the poses of occluded people are almost the same. Earlier techniques that used 2D image-level features would get confused and treat multiple detected poses as belonging to the same person. Hence, the end result would have only one person whose 3D pose is estimated by earlier methods.
With CenterHMR, it detects both players in the above image even though their poses are similar and heavily occluded. It works better with this method because it uses a bottom-up approach for detecting separate people. This means that for each pixel of a body part detected in the image, it tries to assign a point indicating the center of mass. This makes it easier to separate out multiple occluded people using their center points even if they highly overlap.
Thus, it gives us a 3D pose estimation technique that gets us one step closer to using it for real-life applications like analyzing and collecting motion data from sports videos. Once we have this data collection process in place, it will have countless applications in digital game development as well as many other industries.
Useful Links
Thank you for reading. If you liked this article, you may follow more of my work on Medium, GitHub, or subscribe to my YouTube channel.
