SR300 : what is sub-pixel resolution ?

[ABouv]

Hello all,

I am writing my master's thesis and I need to thouroughly describe the SR300.

I do not understand the term "sub-pixel resolution" found in the datasheet, table 4-1, footnote (1) : "The effective range of the camera is up to 1.5m, but the 16b UINT depth format is interpolated over an 8m range (or 1/8mm sub-pixel resolution)."

I have researched sub-pixel resolution, but I could not find anything linking it to depth range.

So my question is : what is the sub-pixel resolution in this particular case ? Why is it equal to 1/8 mm ? And why is the depth data interpolated over an 8 m range ?

Thank you for your help.

[MartyG]

The paper linked to below has some information on depth in structured-light cameras and an explanation of the 1/8 figure. It's an offline PDF that will appear as a file at the bottom of your browser to click on to open. The depth section is on page 51 of the document.

http://bit.ly/2uk1Kr3 http://bit.ly/2uk1Kr3

[ABouv]

Hello MartyG,

Thank you for your answer.

Unfortunately, I previously found this reference and it did not answer my question. I can see why the "sub-pixel refinement process with interpolation factor 8" would lead to a disparity resolution of 1/8 mm, but then I face another problem : what is the link between that statement and the "8 m range" mentionned in the datasheet ?

[MartyG]

I did some further research into your question, The interpolation range seems to be related to the distance that the camera can physically see depth, but not necessarily produce useful data at that depth. For example, it was said that the Kinect V2 camera could see 8m physically with its depth sensor and so had an interpolation factor of 1/8. However in practice, after about 4.5 m the user's joints could not be seen by the camera properly and so useless data was produced beyond that 4.5 m distance.

So it's possible that the RealSense data sheet is saying that in theory the camera could see 8m of depth (hence 1/8 interpolation factor), but in practice the range that produces usable data is much shorter. I may be wrong, but until a better explanation is provided by an Intel representative, that is the best answer I can give.

[ABouv]

That was also my impression. Thank you for trying !

[idata]

Hi Agronymous,

Thanks for your interest in the Intel RealSense Technology. I was looking at this thread and I would like to add the following in regards to your questions. The resolution of the 16bit UINT datatype for Depth corresponds to 1/8 mm. When 1/8mm is multiplied by 65535, which is the highest value of the 16bit UINT, you get the equivalent of 8m. The camera does not really see this far so the effective range is 1.5m or less. In other words, any depth values over 12000 (16bit UINT) should be suspect. Hope this information helps. Regards,-Yermi A.

[MartyG]

Thanks muchly Yermi. I had seen a much simpler version of this during my research into the question which just said '16 bit = 8m', which didn't provide sufficient detail to make it understandable. Your version of the explanation is far clearer. Appreciate it!

[ABouv]

Thank you Yermi.

So if I were to split hair, I would be right in arguing that the exact value is 1/8 * 2^16 = 8192 mm ?

I assumed it was something like that but could not find hard evidence, so thank you.

Now I must ask, for the sake of being thorough : what would you call this "1/8" factor ? Is it indeed the sub-pixel resolution ? And is it correct to use it in the equation of resolution with respect to measured depth (Eq. 2.6) that appears in MartyG's first answer to this thread ?

I also just found today a paper of Guidi et al. (2016) with a possible explanation, showcasing the same equation. Here is the interesting part :

Would you say that we are all talking about the same thing ? Or is "your" 1/8 different from "their" 1/8, in this case delta p ?

Thank you again for taking the time to answer !

Regards,

Agronymous

References :

Blais, F., Rioux, M. & Beraldin, J.-A., 1988. Practical Considerations For A Design Of A High Precision 3-D Laser Scanner System. In R. J. Bieringer & K. G. Harding, eds. Proceeding of SPIE. pp. 225–246. Available at: http://proceedings.spiedigitallibrary.org/proceeding.aspx?articleid=1253256.

Guidi, G., Gonizzi, S. & Micoli, L., 2016. 3D capturing performances of low-cost range sensors for mass-market applications. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives, 41(July), pp.33–40.

[idata]

Hi Agronymous,

The 1/8 factor is just called the sub-pixel resolution. However, we cannot guarantee if our sub-pixel resolution is the same delta p as in the equation. Regards,-Yermi A.