Revision as of 23:18, 13 June 2019

The following section describes some relevant performance measurements during video stabilization for a few standard resolutions. It begins with a summary of the results detailed on the following sections, as well as the pipelines used to capture such results.

Summary

Benchmarking

Standalone element

Pipeline

All the measurements bellow where made using the following minimal pipeline.

PATH_TO_FILE=path_to_file.rgba
WIDTH=1920
HEIGHT=1080
FRAME_SIZE=$(($WIDTH * $HEIGHT * 4))

gst-launch-1.0  filesrc location=$PATH_TO_FILE blocksize=$FRAME_SIZE ! videoparse format=rgba width=$WIDTH height=$HEIGHT ! nvstabilize ! fakesink sync=true

CPU

Resolution	CPU 0	CPU 1	CPU 2	CPU 3	Average (reported by top)
640x480@25Fps	68.48	0	0	56	45.9
1280x720@25Fps	40.97	0	0	36.97	54.5
1920x1080@25Fps	42.05	1.44	1.27	4.055	53,9
2048×1080@25Fps	15.45	0	0	37.3	52,7

GPU

Resolution	GPU
640x480@25Fps	21.58
1280x720@25Fps	29.41
1920x1080@25Fps	37.69
2048×1080@25Fps	35.11

RAM

Resolution	RAM
640x480@25Fps	1442/7861MB
1280x720@25Fps	1443/7861MB
1920x1080@25Fps	1446/7861MB
2048×1080@25Fps	1480/7861MB

RAM consumption for different resolutions.

Glass-to-glass latency

Latency was measured as shown in the picture, and the following are the approximate results. It is important to notice that latency can be affected by the parameters selected on the pipeline, specifically the queue-size property of the gst-nvstabilize element, as further explained on the User Guide. The pipeline used was the following:

gst-launch-1.0 v4l2src device=/dev/video0 ! 'video/x-raw,width=1280,height=720' ! nvvidconv ! 'video/x-raw(memory:NVMM),width=1280,height=720,format=RGBA'  ! nvvidconv ! queue max-size-buffers=1 leaky=downstream ! nvstabilize queue-size=5 ! nvvidconv ! nvoverlaysink sync=false

1280x720@30Fps

queue-size	Latency (ms)
No stabilizer	140.8
1	254
2	258
3	288
4	328.8
5	412.6
6	412

Latency measurements varying queue size.

Previous: Performance

Index

Next: Performance/Xavier

@@ Line 70: / Line 70: @@
 |}
 [[File:RAM tx2.png|frame|center|RAM consumption for different resolutions.]]
-== Full pipeline ==
-The following measurements were taken using an OV10635 camera at '''1280x720@30FPS'''. As further explained on the [https://developer.ridgerun.com/wiki/index.php?title=GStreamer_Video_Stabilizer_for_NVIDIA_Jetson_Boards/User_Guide User Guide], the resulting stabilization, latency and resources the element consumes are all affected by the chosen queue-size, as shown in the results reported as follows:
-<syntaxhighlight lang=bash>
-gst-launch-1.0 v4l2src device=/dev/video0 ! 'video/x-raw,width=1280,height=720' ! nvvidconv ! 'video/x-raw(memory:NVMM),width=1280,height=720,format=RGBA'  ! nvvidconv ! queue max-size-buffers=1 leaky=downstream ! nvstabilize queue-size=5 ! fakesink
-</syntaxhighlight>
-{| class="wikitable"
-|-
-! Queue-size !! RAM !! CPU 0 !! CPU 1 !! CPU 2 !! CPU 3 !! GPU !! Framerate
-|-
-| No stabilization ||  1384/15828MB || 18.875 || 8.625 || 11 || 0.75 || 2.125 || 30
-|-
-| 1|| 1396/15828MB  || 31.11 || 29 || 1.55 || 3.22 || 10.33 || 30
-|-
-| 2 || 1406/15828MB || 23.625 || 22.75 || 9.125 || 9 || 13.125 || 30
-|-
-| 3 || 1426/15828MB || 17.625 || 20.375 || 11.75 || 12.75 || 15.25 || 30
-|-
-| 4 || 1441/15828MB || 16.42 || 20.71 || 14.85 || 14.28 || 8.57 || 30
-|-
-| 5 || 1551/15828MB || 10 || 11.7 || 23 || 11.7 || 7.8 || 30
-|-
-| 6 || 1563/15828MB || 14.77 || 13.77 || 21.11 || 11.77 || 10 || 30
-|}
 == Glass-to-glass latency ==