TDA2PXEVM: Frame Rate for Object Detection and Semantic Segmentation using TIDL

Vijay Krishnan

Part Number: TDA2PXEVM

Hello Friends,

I have implemented Object Detection and Semantic Segmentation using TIDL and using JDetNet SSD.I find the frame rate is on the lower side.For Object Detection I am getting 8.5 fps and for Semantic Segmentation I am getting 1.5 fps.

Are there ways to improve the frame rate?

I have thought about the following methods:

• Divide the Algorithm Link into many light weight algorithm links in order to maintain the fps numbers maintained by other links in the chain.

• Choice of Deep Learning network

• Adjust the InWidth and InHeight used for the DL algorithm to achieve optimum performance

• Change the sync threshold parameters to achieve optimum performance

Please validate them and also please let me know if there are additional methods to improve output fps.(Please note that I am using TDA2P which has 2 EVE cores)

Best Regards,

Vijay

over 7 years ago

0 kumar.desappan over 7 years ago

TI__Mastermind 22145 points

Were you abale to get the expeted performance with the default demo avaialable in the SDk package? If yes, what is fps your observing
Can you share the diffreneces in you trained network compared to TI demo networ?

0 Vijay Krishnan over 7 years ago in reply to kumar.desappan

Intellectual 880 points

Hello Kumar,

With the default demo network I was able to get 8.5 fps for OD and 1.5 fps for SemSeg.With my trained network (JSeg21) I could get 9 fps for OD and 2.5 fps for SemSeg.The expected fps for OD is 22 fps.It would be helpful if you could let me know the ways to optimize and improve the fps for OD and SemSeg.

Best Regards,

Vijay

0 kumar.desappan over 7 years ago in reply to Vijay Krishnan

TI__Mastermind 22145 points

This may be basues of the memory car that you are using. Please erefer the user guide. This demo requires a high speed memory card.
The default demos in SDK shall run as per the adat sheet

0 Vijay Krishnan over 7 years ago in reply to kumar.desappan

Intellectual 880 points

Hello Kumar,
I am using a SanDisk Ultra SD card with a transfer speed of 80 MB/s.Please suggest the required transfer speed.

Best Regards,
Vijay

0 Anuj Gupta over 7 years ago in reply to Vijay Krishnan

TI__Genius 15985 points

Hi,

Can you check at which speed null Src is reading the data?

Regards,
Anuj

0 Vijay Krishnan over 7 years ago in reply to Anuj Gupta

Intellectual 880 points

Hello Anuj,
The Null Src is reading data at the rate of 24 FPS.

Best Regards,
Vijay

0 Anuj Gupta over 7 years ago in reply to Vijay Krishnan

TI__Genius 15985 points

HI,

22 FPS is expected with TDA2x which has 4 EVEs .
WIth TDA2px you should get around 11 FPS.
Can you confirm below ques?

Are you using OPP_HIGH MLO and release binary?
Please use ti provided net and param file for TIDL.

Regards,
Anuj

0 Vijay Krishnan over 7 years ago in reply to Anuj Gupta

Intellectual 880 points

Hello Anuj,
I am using OPP_PLUS MLO and release binary.I am hereby attaching the performance statistics with reduced pipeline for 2 EVEs.I was able to get around16.9 fps.I would like to do further optimizations.The VPE link is having the major frame drop.Can we bypass the RGB conversion and scaling done by the VPE link and achieve higher frame rate?Harman_Performance_Reduced_Pipeline.docx

Best Regards,
Vijay

0 Anuj Gupta over 7 years ago in reply to Vijay Krishnan

TI__Genius 15985 points

Hi,

Tidl expect input in BGR format so you need to have color conversion.
Decoder output data have a larger pitch compared to its width so VPE is also required to crop that additional pich.
Can you check the clock rate of EVE and DSP.
After pressing 0 to close the usecase , it will show the clock for all cores.

Regards,
Anuj

0 Vijay Krishnan over 7 years ago in reply to Anuj Gupta

Intellectual 880 points

Hello Anuj,
Please find the clock values of EVE and DSP given below.They are running at their max frequency.
[IPU1-0] 68.135149 s: #### EVE CLK = 900.0 Mhz
[IPU1-0] 68.135271 s: #### DSP CLK = 1000.0 Mhz
[IPU1-0] 68.135393 s: #### IPU CLK = 212.800000 Mhz
[IPU1-0] 68.135484 s: #### A15 CLK = 1500.0 Mhz

Can we optimize on the pitch parameters of the decode link in order to reduce the dependence of VPE link to do the scaling and cropping?We can optimize the number of buffers used for the scaling operation subsequently increasing the frame rate.Please give your views and suggestions.

Best Regards,
Vijay

0 Anuj Gupta over 7 years ago in reply to Vijay Krishnan

TI__Genius 15985 points

Hi,

You are already getting the better FPS compared to TI demo as you are using OPP_PLUS instead of OPP_HIGH.
You can not remove VPE link as those pitch parameter will come from decoder.
You can try 1 experiment
create a new usecase where you will dump the data of VPE link and then directly feed that data using NUll Src link to OD usecase. So here you do not need to use decoder and VPE link.
But i am not sure whether IPU will be able to read YUV data at that higher FPS.

Regards,
Anuj

0 Anuj Gupta over 7 years ago in reply to Anuj Gupta

TI__Genius 15985 points

Hi,

I haven't heard back from you, I'm assuming you were able to resolve your issue.
If not, just post a reply below (or create a new thread if the thread has locked due to time-out).

Regards,
Anuj

Processors

Processors forum

TDA2PXEVM: Frame Rate for Object Detection and Semantic Segmentation using TIDL