Computer Vision News - June 2020

3 Summary 3D Poses in the Wild 7 generalize in-the-wild images. That’s something that one could explore.” Gerard tells us they have given a lot of thought to what the right metrics are, and he would suggest people don’t focus on one metric alone. Rather, they should be analytical about where the algorithm works and where it doesn’t. He says the team thought one challenge for 3DPW would be to track a person over time, but a lot of people assume they know who they are tracking by placing a bounding box on the person, which in itself can be a challenge. To solve this, the competition has been split into two tracks. One focuses only on the challenge of estimating the 3D pose , while the other focuses also on identifying a person and tracking this person over time . He also says it’s important to look at the tails of the errors – when does the method fail completely? “For example, Michael was tellingme that if you predict an average and you align the overall orientation to the ground truth, you get a pretty good result,” Gerard explains. “This is concerning, right? This means that you can’t only look at one metric, you have to look also at other metrics where it doesn’t, which algorithm is good for what, and why.” We ask if GANs would help to make the challenge more competitive. The team are torn on this. Angjoo is a research scientist at Google and will soon be starting as an Assistant Professor at UC Berkeley . Our readers might remember Angjoo featured as Woman in Computer Vision at ECCV 2018 . She tells us they have explored this direction, but GANs are just one tool in the box. “When there is no ground truth, GANs are one way to enforce a prior knowledge of what human bodies are like,” she says. “You have a real image and a fake image. In this case, the real images are 3D poses captured in the lab and the fake images are the poses that the networks are suggesting. Certainly, there’s a way to use it, but the community has also looked at other ways of enforcing priors, such as VAEs from Michael’s lab.” “Also, we explored priors for pose,” Gerard adds. “In several experiments, what we ended up finding out is that simpler priors do the same job as a GAN.