Computer Vision News - June 2020

2 Summary Co puter Vision Challenge 6 What advice would the team give to people looking to take part? “Don’t rely on training your methods with in- the-lab data,” Michael replies. “We’ve seen again and again that if you just rely on datasets like HumanEva and Human3.6M, they don’t generalize to these more complex scenarios. They’re useful datasets, but if you want to be successful, you’re going to have to find a way to train your method on challenging in-the-wild imagery.” The problem here is the dearth of training data with ground truth 3D pose, so the task for the research community is to figure out how to use real images in complex scenes and train their methods without having access to 3D ground truth. That’s where the team hope participants will bring some innovations or new ideas. “Trying to extract away a lot of information that might not be directly relevant for the pose is a good idea,” Gerard adds. “It’s segmentation and then lifting. The chances are that the network will generalise much better. You have to be careful though because if you remove too much information, then you’re missing out, but there’s some trade-off there.” Aymen, who is a PhD student at the Max Planck Institute for Informatics and supervised by Gerard, agrees: “We’ve seen that if you train your models on indoor datasets but you abstract away information, like if you use optical flow instead of the RGB imagery itself as the input for your network, then the domain shift between the indoor images and outdoor images isn’t so great and it’s much easier for your network to

RkJQdWJsaXNoZXIy NTc3NzU=