Slashdot Mirror
Dell Venue 8 7000, "World's Thinnest Tablet" With Intel Moorefield Atom Reviewed
MojoKid (1002251) writes "Dell recently launched their Android-based Venue 8 7000 slate, claiming it's the "world's thinnest" tablet. It measures a mere 6 millimeters thick, or 0.24 inches and change. That's 0.1mm slimmer than Apple's iPad Air 2 and 1.5mm flatter than the iPad mini 3, giving Dell full bragging rights, even if by a hair. Dell also opted for an Intel Atom Z3580 processor under the hood, clocked at up to 2.3GHz. This quad-core part is built on Intel's 22nm Moorefield microarchitecture. Compared to its Bay Trail predecessor, Moorefield comes in a smaller package with superior thermal attributes, as well as better graphics performance, courtesy of its PowerVR G6430 graphics core. The Venue 8 7000 also features one of the best 8-inch OLED displays on the market, with edge-to-edge glass and a 2560x1600 resolution. Finally, the Venue 8 7000 is also the first to integrate Intel's RealSense Snapshot Depth Camera, which offers interesting re-focusing and stereoscopic effects, with potentially other, more interesting use cases down the road. Performance-wise, the Venue 8 7000 is solid enough though not a speedster, putting out metrics in the benchmarks that place it in the middle of the pack of premium tablets on the market currently."
Okay, actually, 100 microns (0.1mm) is a reasonable diameter for a human hair. So, kudos for the phrasing!
The reviewer should be embarrassed, and so should you for not reading up on RealSense, but it's probably unintentional.
The error is because stereo depth accuracy is quadratic, it degrades as the square of the distance to the sensors. The distance (baseline) between the cameras in a RealSense unit is so small that any distance measured beyond a few metres is inaccurate. It was a stupid thing to demonstrate, but it shows that many reviewers (and users it seems) don't understand the limitations of 3D measurement systems. For this reason, Intel clearly states that RealSense is only good up to 10m (and even then I would be sceptical that it works well beyond 5).
This is easily verifiable with your eyes. As an object gets further away, it becomes harder and harder to determine its distance because the optical parallax of the object tends to zero (i.e. it appears in the same x-position on each of your 'sensors'). Try it next time you're in a car or on a train, we all know that nearby objects appear to whizz past while background features like mountains/hills remain stationary.
Specifically the error equation is dZ = Z^2/bf (the distance measurement is is Z = bf/d where d is the disparity (parallax) in pixels)
Where dZ is the distance error, Z is the target distance, b is the baseline and f is the focal length in pixels. I've assumed that you can detect correspondences to within one pixel, realistically it'll be better than that for a competent stereo matching algorithm. Now in this case Z is several hundred metres, b is order 100mm and f order 1000px.
Do the maths: 100^2/(100e-3 * 1000) = around 100m error. At 5m? It's around 25cm and 1m it's 1mm. The actual numbers will be different because I don't know the exact baseline, or the focal length. I can tell you for sure that the cameras aren't high enough resolution for that to make a significant difference to the accuracy.