TDA4VM: DOF result have a large deviation.

chao yang

Part Number: TDA4VM

Hi,Fready,
Title:
DOF result have a large deviation.
Purpose:
After many tests, we find the results obtained by our 7.3SDK DOF demo have a large error.
while tracking results of Farneback in OpenCV are basically consistent with our LK optical flow.
We hope to get similar results to our LK optical flow based on DOF demo code.
Reference:
1.Our demo program and result could be seen by attachment: DOF.7z, Result_Image.7z.
2.Mark methods:
(1) Take corner points;
(2) OpenCV code could be seen by attachment: Tda4vsopencv_opticalflowfarneback.7z (OpenCV version: opencv_python-3.3.0.9-cp36-cp36m-win32, Python: Python-3.6.2), the video will be sent to you Wechat;
(3) DOF marks (u, V) of corresponding corner points in bin files, and Matlab tracing points is enough;

/cfs-file/__key/communityserver-discussions-components-files/791/wan.r.zip

over 4 years ago

0 Hrushikesh Tukaram Garud over 4 years ago

TI__Intellectual 2615 points

Hello Chao,

I will take a look at your question. In the mean time, can you please give me following details-

As I understand you are observing quality difference between version of LK you are using and the DOF output. Also, you are interested in sparse output. Is this understanding correct?

Are you trying to run DOF demo on EVM?

Thanks and Regards

Hrushikesh

0 Hrushikesh Tukaram Garud over 4 years ago

TI__Intellectual 2615 points

Hello Chao,

I have gone through the point-pair plot image you have shared. My observation is that the flow magnitude is higher for TI DOF estimates. The algorithms being different than OpenCV, I would expect marginal difference in the output. However, I can not comment on the accuracy of either OpenCV or DOF results as I do not have access to the original input images or the ground-truth.

Recently, there have been some changes to the DOF application. I will transfer your query to the concerned so you can migrate to the updated version if necessary. They will also be able to check the other details you have shared in DOF.7z.

Thanks and Regards

Hrushikesh

0 Brijesh Jadav over 4 years ago in reply to Hrushikesh Tukaram Garud

TI__Guru**** 501525 points

Hi Chao,

Can you please use SDK8.1 for checking this out? We have done several changes in the DOF demo, so could you please check this in SDK8.1?

Regards,

Brijesh

0 chao yang over 4 years ago in reply to Hrushikesh Tukaram Garud

Prodigy 170 points

The original image we have emailed to Fredy.

From our experimental results on SDK 7.3, it can be seen that there are errors in the DOF results, such as the movement trend, not just the problem of precision.

Please tracking in our image using your method.

0 chao yang over 4 years ago in reply to Hrushikesh Tukaram Garud

Prodigy 170 points

Hello, Hrushikesh

We used LK DOF OpenCV_farneback three methods to track the same video and take a frame of image comparison results.

We hope to use DOF to get results as precision of LK or OpenCV。

Our result is on EVM.

0 chao yang over 4 years ago in reply to Brijesh Jadav

Prodigy 170 points

Hello,Brijesh

What change in SDK 8.1 for DOF? SDK8.1 has many problems now and cannot do DOF experiments.

0 Brijesh Jadav over 4 years ago in reply to chao yang

TI__Guru**** 501525 points

Hello Chao yang,

What problems are you seeing in the SDK8.1? SDK8.1 works fine.

DOF example in the SDK8.1 is more fine tuned and is matching with the PC emulation output, so could you please try it out?

Regards,

Brijesh

0 chao yang over 4 years ago in reply to Brijesh Jadav

Prodigy 170 points

Hello,Brijesh

We are adapting sdk 8.1.Currently we focus on the use of 8.0 and try as soon as possible on 8.1. Can you first explain what improvements 8.1 has made in DOF? Can you provide documentation about DOF in 8.1?

0 Brijesh Jadav over 4 years ago in reply to chao yang

TI__Guru**** 501525 points

Hello Chao yang,

As i mentioned, we have changed confidence tree in the latest release and also fixed few issues/bugs in the SDK8.1 release. So could you please check it in the SDK8.1?

Regards,

Brijesh

Processors

Processors forum

TDA4VM: DOF result have a large deviation.

Our result is on EVM.