At this time, we (the developers) don't have WebCam Pro 9000 in our test lab because they don't sell it in Moscow as of yet. Logitech claims WebCam Pro 9000 is a minor revision of QuickCam Pro 9000. Logitech never discloses official information about their cameras FOV, so we are not 100% sure it did not change.

You can help us investigate this problem by capturing camera calibration images for WebCam Pro 9000 for iPi Camera Calibration Tool and sending them to me:

http://www.ipisoft.com/iPiCalibrationTool.php

I would appreciate if you could capture calibration image sets (a set should be at least 3 images) in different resolutions (320x240, 640x480, 800x600) as Logitech cameras are known to have some strange unusual image distortions depending on capture resolution.

http://www.downloadtaxi.com/d/1259265367

here are the webcampro calibration pictures.

By the way, how's Moscow? Is it snowing there already? Does it snow there in during these months?

You can ignore those questions. :)

Thanks!

There is noticeable distortion of the pattern in pictures. Is it camera distortion, or paper crumpling? (Don't get me wrong, I am not being picky, this is just important for calibration)



Very cold and it was snowing a couple of days ago. We have 5 months of winter in Moscow. That's why USB-heated gloves are popular here: http://mobbit.info/item/2005/11/15/usb-perchatki

Distortion? I am not too sure but I used photopaper for the print so the paper's not crumpled and disabled auto focus on my webcampro before I took these pictures. Well, I used rightlight to brighten it up a bit. And I tilted the print 45 to 10 degrees approximately as it was explained in the Calibration tool. The result was 77.29453 in the 320x240 size.




Oh that's stone age technology man. Here in america, we can download our own brain memories to a USB, but that's only to the common folks around here.

Actually, I am scheduled to get some surgical enhancements to finally plug myself directly to the internet as you can see the example image at the right. Santa picked it up from my wish list because I have been a good boy the past year so I will get it this Christmas.

Thanks for that, I'm glad to hear I'm doing it right :)



The rough positioning isn't that arduous so I'll probably do that every time. The results I got from my calibration were:
Cam1 to Cam2: 3.66m
Cam1 height: 1.50
Cam2 height: 1.93
Cam3 height: 1.54
Which are pretty similar to your results, between 2 and 10cm different.

While my video started in the dark, in retrospect I think it makes sense to record a T-Pose before turning the lights out to calibrate. Then you can a) line up the virtual cameras onto the image of the real cameras - a real help with the rough manual alignment, and b) compare your actor to the model to further improve the manual alignment.

-Ian

That's right.

I hadn't thought of using the T-pose to start the calibration video. That's an excellent idea.

That said, I am getting good results with my current calibration right now. The only trouble I have is (1) one arm sometimes gets "lost" -- it ends up behind me or at a funny angle; and (2) If I sit or lie on the floor, I don't get good results at all. Anything I can do about either of those two problems?

There can be different reasons for this and different strategies for fixing this.

1) It is possible that arm was occluded. Using 4 cameras can help to mitigate this problem. For tracking to work correctly, it is preferable that each arm should be visible in at least 2 cameras all the time.
2) It is possible that arm position was recognized incorrectly during pose transition. Transitions usually take small time - like 2 or 4 frames. In this case, you can manually correct the pose in few problematic frames and repeat tracking.
3) There could be a problem with dim lighting or background clutter, so that software mistakes actor's arm for part of background. In this case, it is recommended that you improve lighting or background and re-shoot the video, because manually editing many frames can be too time consuming.



We currently research this problem. Preliminary I can say that it can be related to placement of cameras - tracking sitting or laying actor is not working well with traditional placement of cameras (one center, two on sides). It probably should work better with 4 cameras. Future versions of our software may improve on this.

Perhaps a very high angle on the cameras would be better for poses lying down, so it almost looks like you're standing up?

Good info with the occluded arms. I'm going to be doing a lot of hand gestures and practically nothing behind the body, so I might position my left and right cameras within about 60-70deg of the front cam, instead of 90deg. That way all three cameras should be able to see the hands at all times, even when they are by the actors side.

-Ian

In theory, yes.