Computer Vision News - June 2020

3 Summary 3D Poses in the Wild 5 and joints were computed in all of the images. However, there were only a small number of subjects, the variety of clothing was minimal, and the cameras weren’t moving. It was very much a laboratory environment. A few years later, another dataset, Human3.6M , came along. It was bigger than HumanEva, but again based in a laboratory environment, with limited clothing and a limited number of subjects and motions. There was no camera motion. The 3D Poses in the Wild (3DPW) dataset is the next generation. It was captured in the wild, outside of a motion capture lab, and combines various sensors to give accurate reference poses. So far, people have been evaluating the dataset using different protocols and metrics, which can be difficult to compare. The purpose of the 3DPW Challenge is to standardize this so that researchers compare their methods in a consistent manner in future. The team will evaluate the challenge according to several clearly organized metrics. Gerard is the head of the Real Virtual Humans research group at the Max Planck Institute for Informatics . His lab has posted the dataset, which has ground truth and includes elements that were missing from previous datasets, like multiple people in a scene all interacting and occluding each other. This presents challenges for the current algorithms. Compared to other object classes, humans see a lot of variability in identity, shape, articulation, clothing, and appearance. Most algorithms work well on a single, well-isolated image of a person, but when there are lots of people interacting, they can fall down. The team are keen to point out that this is not simply a detection challenge. The challenge here is to estimate the 3D articulated pose of the body . That’s more complex than a detection problem or even a rigid object pose estimation problem where you might have six degrees of freedom. Here, there may be 30 degrees of freedom that you need to estimate, and parts of the body may be occluded – for example, if a person is seen from the side, you may not see one of their arms.