Bandwidth consumption of quantized/compressed depth images (SIM)

[DatSpace]

Description

Depth images are represented with an array of float numbers. Float is 32 bit and allows for up to 7-8 significant digits.

The camera currently used (Realsense D455) has a theoretical accuracy on the z axis of <2% at 4m, which makes it to have an error of 8cm at worst, at 4m.

In other words, the accuracy of the realsense is at best in the centimeter scale and all of the other significant digits in the float can be discarded. Normally that wouldn't have any effect on the data themselves as they are still represented in a float, unless you do lossless compression (as we do with zlib).

Data

For the experiment we try different quantization values (significant digits) for a depth image from gazebo, while the robot is looking turned 45 degrees at a wall 1m away from the camera. The bandwidth bellow is per frame averaged over 10 seconds, transmitted through LAN with zlib compression.

Significant Digits	Bandwidth	Percentage Change	Comments
8	7.23 Mbps	-
7	7.35 Mbps	+1.66%	It appears that at that many digits it has no effect and it just seems random
6	7.04 Mbps	-4.2%
5	6.05 Mbps	-14.0%
4	1.77 Mbps	-70.7%
3	319.2 kbps	-82.4%	Millimeter accuracy
2	112.8 kbps	-64.6%	Centimeter accuracy. Seems to be the most logical for our use case, as the sensor it self is limited to that.
1	101.6 kbps	-9.9%	Decimeter accuracy with visible quantitation in the depth image color.
0	71.44 kbps	-29.7%	This represents only integer values in meters and obviously cannot be used.

Note: There was some background data transfer at the range of 30kbps that is significant for the lowest measurements. Still the numbers are so small that it doesn't really matter.