3 ms·
Researchers at Adobe published similar work several months ago: https://news.ycombinator.com/item?id=20978055 https://news.ycombinator.com/item?id=20978055 In
by Mr_P 7y ago
Researchers at Adobe published similar work several months ago:
https://news.ycombinator.com/item?id=20978055 https://news.ycombinator.com/item?id=20978055
Interestingly, FB folks used a vastly simpler network architecture, but it probably still works because they have access to so much real data. In contrast, the Adobe paper had a fairly involved network with pretrained VGG features, possibly to compensate for only having synthetic data to train on.
- dheera 7y agoIt's also worth noting that FB doesn't need a lot of accuracy in the depth data for this to be "good enough" for a simple wobble effect on a phone. It definitely doesn't need to be held to the same standards that would be held for a monocular depth estimator for, say, an autonomous vehicle or robot; in FB's case it doesn't matter whether the tree is 4m or 6m away, and really only matters that it's roughly behind the dog or cat or whatever. Separately, depth estimators like this are also useful to fake the depth of field effects of large-aperture lenses, which I'm pretty sure is used on the latest Pixel and iPhone in their background-blurring "Portrait" mode; Facebook could quite possibly include that as an in-app feature to make such effects accessible to everyone with cheaper phones as well.