Best camera for scanning faces (ear to ear)

[AJage]

Hi there, I have been lurking around the forum for a couple of days now. And I have come to the conclusion that for scanning faces the best options are the SR300 and the D415.

My requirements are that the scanning must be done as accurate as possible while still being able to scan fast. Because the usecase would not be in ideal circumstances where people are able to sit down for more than 10 seconds.

The faces will be nearby, (0,3-1,0m).

But I'm torn between the two because of a couple of things.

1. The SR300 is now discontinued and will probably not receive considerable updates anymore.

2. The D415 has not been implented yet for my usecase (scanning faces).

Now, what camera would be better to use for my usecase? Will the SR300 provide better accuracy in this short distance, or will the D415? And/or will the D415 be able to scan faster, or will the D415?

Hopefully, someone can help me with these questions.

Thanks in advance,

Arnand

[MartyG]

When you say you want to scan faces, I am guessing that you want to make a 3D model scan of the face. Is that correct, please?

If you need to make a rapid head scan of the face without having to do additional editing processes afterwards, the R200 model is actually likely to be the best choice. This is because:

1. It is well-supported by commercial 3D model scanning software

2. The 'RealSense SDK 2016 R2' SDK that is compatible with the R200 comes with a free sample program called 3DScan that can scan heads and then output the scan as an .obj format 3D model file.

3. It is especially suited to scanning large objects and the human body, unlike the SR300 which is suited to scanning small objects.

Whilst is is possible to make 3D scans with the 400 Series cameras, it is currently a slower process because you have to first take a 'point cloud' scan made up of dots and then use another software package such as MeshLab to convert the point cloud into a solid model.

The R200 is also a retired model, but is still available from Intel's online store in a heavily discounted bundle deal with a single-board computer, the Up Board.

https://click.intel.com/intelr-realsensetm-robotic-development-kit-2561.html Intel® RealSense™ Robotic Development Kit - Intel® RealSense™ Depth Cameras

If the D415 is a preferable option, you could minimize the time needed for people to be seated by taking the initial point cloud scan and then converting the saved point cloud file to a solid model later on at your convenience. The RealSense Viewer software that comes with the D415's RealSense SDK 2.0 can create point cloud files quickly.

[AJage]

Thanks for the quick reply.

Sorry, I was not clear enough. I indeed meant on creating a 3D model, with and without texture.

Regarding the R200, does'nt it greatly compromise on accuracy?

Also, the point cloud, received from the D415 does only export from one angle, does'nt it? It does not support tracking yet?

I did find 3D-scanning software that recently supports the D400 series: RecFusion. (€99, but evaluation version is free).

About the SR300, it is also supported for the Easy3DScan app, I recall. So it would be a better choice? (If I could get a unit)

Greetings,

Arnand

[MartyG]

The old RealSense models like the R200 have an error range of around +- 5% and the 400 Series have an error range of 1% or less. I wouldn't say that the R200 is bad at scanning, though I have sometimes seen reports where the dimensions of a scan are off by a few centimeters compared to the original.

It is true of any 3D scan with a single camera that you are going to either have to rotate the camera or the person if you want to get a 360 degree scan of the head instead of just a frontal scan.

If you are interested in skeletal tracking for the 400 Series cameras, I recommend checking out the Nuitrack software.

https://www.youtube.com/watch?v=gMPtV4NXtUo Intel RealSense D415/D435 and Nuitrack skeletal tracking SDK replace Kinect SDK - YouTube

If you cannot get an Intel SR300, the next best choice is the Creative BlasterX Senz3D. This is fully compatible with the RealSense SDK 2016 R2 and 2016 R3, and should work with the new SDK 2.0 that the D415 / D435 use (as SDK 2.0 supports SR300 too). This is because Creative manufactured the official RealSense SR300 camera for Intel, so the BlasterX is virtually identical. Another choice f BlasterX is unavailable is the Razer Stargazer, another SR300-compatible camera.

[SGhos50]

Hello,

We are also trying to use D-415 to scan faces from ear to ear. In our setup D-415 is fixed and the closest the face is to the camera is 36cm (so our Z-range is somewhere between 36 to 50 cms). I have read the white paper on tuning the depth accuracy (https://www.intel.com/content/dam/support/us/en/documents/emerging-technologies/intel-realsense-technology/BKMs_Tuning_RealSense_D4xx_Cam.pdf) and currently am using the following parameters :

• High Accuracy preset
• Disparity : 30
• Resolution : 1280 x 720 @ 30 fps
• Laser Power 250mW

Our depth accuracy requirements are upto 1mm. Is it possible to achieve this with D-415 using only the IR imagers? Any help would be much appreciated.

Also has anyone tried to compute the deviations of a grid of points in a volume for this camera and would know how to proceed?

Thanks

[MartyG]

Hi @SGhos50​  thanks for your question! RealSense support has now moved to the forum in the link below, so please post any future questions there. Thanks!

https://support.intelrealsense.com/hc/en-us/community/topics

In answer to today's question: the D415 has a larger minimum distance than the D435 model, making it more difficult to get very close to observed surfaces before the image starts breaking up as the camera distance (MinZ) goes below its minimum. You can reduce the minimum distance and get closer by increasing the value of a setting called Disparity Shift. Increasing the Disparity Shift value reduces the minimum distance, allowing the camera to get closer to a surface. Doing so also reduces the maximum observable distance (MaxZ), though this should not be a problem with close-range sensing.

The D415 has around 2x better image quality than the D435 model and a higher optimal depth resolution (1280x720), so it is a good choice for close-range, accurate scanning.

If you are operating at very close range, I also recommend reducing the Depth Unit Scale from its default setting of '0.001' to '0.0001' to help avoid quantization effects.

Intel's excellent camera tuning guide for the 400 Series camera states that "when operating at very close range, the D4xx cameras can inherently deliver

depth resolution well below 1mm". The guide can be found here:

https://www.intel.co.uk/content/www/uk/en/support/articles/000027833/emerging-technologies/intel-realsense-technology.html

For deviation analysis you could try the open-source program CloudCompare, which is compatible with the 400 Series cameras.

https://www.danielgm.net/cc/

[SGhos50]

@MartyG​ thanks for your reply, I have few follow up questions, then I will create the same topic in the new support link you send and followup there.