Hi! bit of a newbie, but i got the Optical Flow SDK compiled and running (here’s an example of some 4k and 1080p footage of local jungle animals Test Optical Flow Sloth and Monkeys - YouTube )
but i am a bit confused when i run the AppofCuda on a series of about 200 pngs it takes about 3 minutes to process
but then it will say “Total frames = 224 time = 1.04254 s, NvOF FPS=53.3993”
which seems far from the truth because it took like 3 minutes to run. What’s going on? This is the release version, not the debug? Is there some option where i can experience the actual speed?
Hi.
In the sample application, ‘time’ and ‘fps’ is measured across NVOF APIs. Other operations like disk IO, uploading data from system memory to video memory etc is not considered in performance measurement.
Said that 3min duration to execute looks too high. What is the command line, GPU, and OS version you used? Do you get the same result consistently across multiple runs? It will help if you provide all the details to reproduce this issue internally.
Is there some option where i can experience the actual speed?
NVOFA programming guide in chapter 8 'Guideline for efficient usage of NVOF API" describes how to tune application to maximize performance.
Thanks.
I am using a Windows 10 HP Omen laptop with an Nvidia RTX 2070 graphics card. Set on max performance mode.
Every time i run, it seems to take about the same amount of time. I don’t have any other programs open that would be using these resources either.
Now that I look at the other files generated , i notice the RAW OF binary and middlebury.flo files are MASSIVE
and so i disabled their output
I ran a version where i disabled the output of the optical flow binary files and just had the RGB visualization generated and it only took 1 minute to finish.
Without ANY output being written to a disk, this same test of the 1080p footage over 200 frames only took 16 seconds.
so yeah, looks like writing to disk is what eats it up :)