Computer Vision News - June 2020

2 Summary Co puter Vision Challenge 4 3D Poses in the Wild Challenge 3D human pose estimation has seen a great deal of progress recently with a number of respected papers being published on the subject. However, results for natural outdoor in-the-wild scenes have been largely qualitative, with quantitative evaluation generally being limited to indoor datasets. To redress this balance, Michael Black, Gerard Pons-Moll, Angjoo Kanazawa and Aymen Mir have organized the 3D Poses in the Wild Challenge. We speak to them about the challenge and work to date on human pose estimation. “The field has been working on these problems for a long time,” Michael tells us. Michael is a founding director at the Max Planck Institute for Intelligent Systems , where he leads the Perceiving Systems department . We first spoke to Michael at ECCV 2016 a nd are honored to welcome him back to Computer Vision News. “When I started working on human motion estimation there was no ground truth at all. There was no way to evaluate your results. Everybody used a new sequence. They showed a result on one sequence and typically they showed the result from the camera view that the thing was taken from. There was no way to compare and you couldn’t know if the field was actually progressing.” Michael led a team that introduced the first human motion and pose estimation dataset, HumanEva , which provided camera capture with a motion capture system synchronized to it. It used a Vicon motion capture system with reflective markers. People wore natural clothing and markers were put on in ways that were not too obtrusive. Video was captured at the same time and ground truth 3D poses and 2D locations