CVPR Daily - Wednesday

10 DAILY CVPR Wednesday Poster Presentation “ A previous work called Bi3D predicted binary in-front/behind planes for stereo networks, ” he explains. “ That is one of the most related works, but the difference is they were trying to predict depths efficiently, within ranges they cared about. We’re predicting the binary in front/behind but on a per-pixel basis. The in-front/behind question – is this pixel visible or not? – is on different 3D positions in the world, so we need to be able to query it differently rather than just an in- front/behind single plane as they did. ” One may wonder why the direct prediction of occlusions has not been explored before. Jamie points out that huge efforts were dedicated to advancing depth estimation techniques , which needed to reach a certain point before the occlusions even looked reasonable. Additionally, academic research has often emphasized the accuracy of benchmarks: How good is the model at predicting depth? How good are the 3D reconstructions? Rather than how good it looks in an AR experience, which is what companies like Niantic care about . “ It was coming from the industry side that made me think, okay, this is a real problem for us at Niantic, ” he says. “ I think it should be a real problem for academia as well, and we should think about it as a task. There should be research into this area. It shouldn’t just be companies thinking, how do we solve it? It should be researchers worldwide thinking, how can we make this look better? ”