At the Embedded Vision Summit, a Microsoft representative and one of the organizers of the event talked about their experience with the HoloLens helmet. Recall, this is an experienced Microsoft helmet for augmented reality on the Intel platform (SoC Cherry Trail). The device is self-contained and does not require a connection to a PC. Part of the task is performed by the helmet platform (Intel Atom processor and specialized HPU / Holographic Processing Unit DSP), and some tasks are performed
in the cloud when transferring data via Wi-Fi. The helmet is already distributed among developers. The issue price is $ 3000. As the expert said, the device should become easier (now the helmet weighs a pound) and is cheaper.
Later in the current year, the release of a lighter helmet at a reduced price is expected. From the details it is known that instead of four cameras, the "discounted" version will have two cameras, and it can be connected to a PC running Windows 10 to achieve the effect of immersion in virtual reality. Also announced is the support of head tracking technology. At the same time, the developer notes that to HoloLens on the mass-scale characteristics become similar to Glass Glass glasses
should pass at least ten years. But even from the current version of the helmet "the head is no longer split."
One of the main challenges faced by the development of HoloLens is the need to process a large amount of visual information. In other words, the problem of creating a digital model of the space surrounding a person. At the moment, Microsoft is looking for a balance between the amount of data processed by the helmet and the amount of data processed in the cloud. For example, to create a 3D map of a four-storey office of one of the developers it took about 100 sessions. The data was "glued
together" in the Microsoft cloud using the resources of remote graphics processors. But even this was not allowed to avoid mistakes in the form of "black holes" on the map and other distortions.
An alternative may be to "glue" the card using processor resources and floating point operations, as well as attracting neural networks to this work. Ideally, the developer is sure, all this work or most of it should be performed directly in the image sensor and this should be sought. Otherwise, it is necessary to chase back and forth large volumes of information. Well, the experiments continue and the company expects the support of the project from independent developers. And
these day by day it becomes more and more. Related Products :
|