none
Hardware Requirements to get Satisfactory Resolution and Reconstruction from a Kinect V2 Scan RRS feed

  • Question

  • I am new to the Kinect for Windows V,2 (KFW2) Sensor, and as a prospective hobbyist  I am finding it difficult to determine the ”best for me” hardware requirements to acquire a satisfactory resolution and reconstruction from a scan to use in a 3D printer.

    Please excuse and correct any mistaken impressions I expose in the following overview of what I have learned so far.

    It seems that the measure of success to achieve my goal is to achieve 30 Frames per Second (fps) using applications such as Fusion Explorer and 3DBuilder while making a acan.

    After watching many video demos I have determined that the stated minimum system requirements published by Microsoft are focused on just getting the KFW2 sensor to actually “work” and have a mesh constructed regardless of quality. Beautiful models have been build using these applications, but in many of them I have heard both explicit and implicit comments that the computer used was much more capable than the stated minimum. In fact, in one such video the Microsoft presenter said words to the effect that “…we have some high-end computers here but you need more than that to get good results in real time...” In many cases a Surface Pro 3 was somewhere in the scene. Well I am using a Surface Pro 3 i7, and I do not get satisfactory results.

    So my question is, “Has anyone achieved a frame rate of 30 fps with Fusion Explorer and/or 3D builder? If so would you mind sharing the make and model of the computer you used, or the specs of the components you used (CPU, GPU, USB 3.0 controller, etc.) if you built your own custom computer.

    Thanks in advance         


    jperry

    Thursday, March 19, 2015 6:41 PM

Answers

  • Hi JPerry,

    I have a laptop with a 4GB 680m and a intel i7-3820qm (USB 3.0 eXtensible) and get a solid 30 fps fusion which takes around 60% GPU load with a 640 vox/m 256x256x512 grid. FPS drops to 17-20 when I increase the numbers to 768 vox/m, 512x512x768 grid.

    On my Surface Pro 3 256GB i5-4300U I can get a decent 20-25 fps with 384 vox/min 128x128x128. 

    Fusion is very memory intensive, hence the frame rate dependency on volume size. I think in the original Fusion paper they mention that each voxel element takes up 32bits of memory so for 512x512x512 it takes up 512MB of GPU memory. 

    I would say anything with 1.5GB of GPU memory, later than a (desktop) nvidia GTX 660 (or AMD equivalent) should get you a solid 30 fps unless you use huge volumes, then I would consider using some of the fusion alternatives such as Voxel Hashing or KinFu Large Scale. 

    http://www.graphics.stanford.edu/~niessner/niessner2013hashing.html

    http://pointclouds.org/documentation/tutorials/using_kinfu_large_scale.php

    Although you may have to do some work getting those working with the v2.

    • Marked as answer by jperry Friday, March 20, 2015 6:10 PM
    • Unmarked as answer by jperry Friday, March 20, 2015 6:10 PM
    • Marked as answer by jperry Friday, March 20, 2015 6:11 PM
    Friday, March 20, 2015 3:18 PM

All replies

  • Hi JPerry,

    I have a laptop with a 4GB 680m and a intel i7-3820qm (USB 3.0 eXtensible) and get a solid 30 fps fusion which takes around 60% GPU load with a 640 vox/m 256x256x512 grid. FPS drops to 17-20 when I increase the numbers to 768 vox/m, 512x512x768 grid.

    On my Surface Pro 3 256GB i5-4300U I can get a decent 20-25 fps with 384 vox/min 128x128x128. 

    Fusion is very memory intensive, hence the frame rate dependency on volume size. I think in the original Fusion paper they mention that each voxel element takes up 32bits of memory so for 512x512x512 it takes up 512MB of GPU memory. 

    I would say anything with 1.5GB of GPU memory, later than a (desktop) nvidia GTX 660 (or AMD equivalent) should get you a solid 30 fps unless you use huge volumes, then I would consider using some of the fusion alternatives such as Voxel Hashing or KinFu Large Scale. 

    http://www.graphics.stanford.edu/~niessner/niessner2013hashing.html

    http://pointclouds.org/documentation/tutorials/using_kinfu_large_scale.php

    Although you may have to do some work getting those working with the v2.

    • Marked as answer by jperry Friday, March 20, 2015 6:10 PM
    • Unmarked as answer by jperry Friday, March 20, 2015 6:10 PM
    • Marked as answer by jperry Friday, March 20, 2015 6:11 PM
    Friday, March 20, 2015 3:18 PM
  • Phil, Thanks. This is extremely helpful.

    jperry

    Friday, March 20, 2015 6:12 PM
  • Hi, Phil Noonan

    Can you tell me more about your " laptop with a 4GB 680m and a intel i7-3820qm (USB 3.0 eXtensible)hardware"

    which Ghz, GPU?

    Wednesday, January 6, 2016 1:16 AM
  • Hi,

    http://ark.intel.com/products/64889/Intel-Core-i7-3820QM-Processor-8M-Cache-up-to-3_70-GHz

    http://www.geforce.co.uk/hardware/notebook-gpus/geforce-gtx-680m/specifications

    Wednesday, January 6, 2016 8:02 AM