Stereo Correspondence Block Matching

This video shows some tests with opencv stereo correspondence block matching algorithm with two cameras. The middle window shows the depth map obtained by the left and right cameras above, in realtime.

Category:

Education

Tags:

• stereo
• opencv
• block matching
• correspondence

[MrCybin]

Hi. I'm trying to use this with two logictech webcams. The result is really poor, not as good as this. Do you have tips for the set up of the stereo rig?

How far apart and at what angles should the cameras be before calibration?

Is individual calibration of the cameras a wise step before stereorectify?

My code works well for the example images that come with OpenCV. Just not with these two cameras as a stereo-rig.

Thanks,

D

[dannyxyz22]

@MrCybin Hi MrCybin. I put the cameras very close (as close as possible, probably 4 or 5 inches). I didn't calibrate the cameras, I just used them straight ahead, but it seems that calibration is a must for good results. Well, I believe that the best solution for depth image nowadays is using Kinect, as my webcam results aren't good neither... :)

Kind regards,

Daniel

[VanDucTech]

Hi, can you show your input Parameters with correspondence block matching algorithm .

Mine is: BMState->preFilterSize=41; BMState->preFilterCap=31; BMState->SADWindowSize=41; BMState->minDisparity=-64; BMState->numberOfDisparities=1­28; BMState->textureThreshold=10; BMState->uniquenessRatio=15;

But it seem not good

Thanks.

[dannyxyz22]

@VanDucTech Hi VanDuc :)

It was made a long time ago and I don't have that back, but I used default OpenCV parameters. One thing I did was using cameras from the same brand and model, and one thing that would enhance the results would be calibrating the cameras :)

Kind regards,

Daniel

[dannyxyz22]

Hi sekomaru,

I'm using highgui for output, for sure. It won't be a problem if you open another window for that. I guess most of this code comes from OpenCV samples, if I'm not mistaken.

I guess better results should come had I calibrated the cameras first.

Best regards,

Daniel

[diabelliforlife]

thanks a lot danny, I read your suggestions and i finnally fix my results, but now I have another problem....you see, I tried to go further on the Learning Opencv's example of the chapter 12 by using the cvRepreojectImageto3D() function, but well, it seems that it only works when you feed the set of images from the sample file... I tried to feed my own set of image pairs but it doesn't work at all....could you please help me???

Thanks in advance!!!

[dannyxyz22]

@diabelliforlife Sorry... haven't tried those... yet :)

[7hanansan]

Yes, This is seems to be Great. If u change the window size(Block Size) , then you can get a better Result. BM is best Algo for Real time App.

[MrCybin]

@MrCybin Thanks for getting back to me. Could i email you some results so you can see what is going wrong?

[sekomaru]

looks pretty good. are you using highgui for your output? i always get a simplified version of the image and have to save them, and then open them up outside of my program to see my output correctly