Feet being clipped

Greenlaw · **Posted:** Wed Aug 22, 2012 11:53 am

Hi,

Still playing around the the 'Three Happy Cats' data in iPi DMC 2.0. When tracking, it seemed like iPi Studio was having a lot of trouble figuring out where the feet were so I took a closer look at the data in Player. Here's what I found:

Attachment:

2012-08-22_111410.jpg [ 211.73 KiB | Viewed 20601 times ]

It looks like her feet are being completely clipped by the ground plane. When she touches the floor, her hand also becomes clipped from the volume data. I was wondering if you've seen anything like this? Could it be that the calibration off because she is standing on a foam mat that is about 1.5 cm thick?

Here are a few more frames to further illustrate the error:

Attachment:

Sequence.jpg [ 1.13 MiB | Viewed 20601 times ]

I haven't compared this with results from iPi DMC 1.0 yet. From what I recall, it's probably the same though.

A while back, I meant to send you guys the original data to look at. Sorry about this delay--I will send you the original iPi 1.0 data today.

G.

Greenlaw · **Posted:** Wed Aug 22, 2012 11:56 am

Although, now that I'm thinking about it, standing on the mat should have made her feet even more visible since it elevates her 1.5 cm.

Very curious.

G.

Greenlaw · **Posted:** Wed Aug 22, 2012 11:58 am

I don't know if this only occurs with 1.0 data that's been converted to 2.0. I'll create an original 2.0 motion test today and let you know what I see.

G.

vmaslov · **Posted:** Thu Aug 23, 2012 9:48 am

Most probably this is the result of background subtraction. With such camera orientation any slim body parts like feet and hands touching the floor have a very small difference in depth with the background (the floor itself). So background subtraction algorithm assumes this is just some noise and replaces it with the background. We also have noticed such unwanted behavior with adult actors when depth sensor is placed considerably high (at 1.5m) and looks down at considerable angle.
This is indeed a problem which has become more obvious now. We need to figure out some improvements to background subtraction algorithm that will eliminate this issue.
However, I can assure you that this is not a novelty of v2 and the tracking quality does not suffer compared to v1. In v1, essentially the same algorithm was used by the tracker under the hood. You just did not see its results.

Greenlaw · **Posted:** Thu Aug 23, 2012 12:42 pm

vmaslov wrote:

However, I can assure you that this is not a novelty of v2 and the tracking quality does not suffer compared to v1. In v1, essentially the same algorithm was used by the tracker under the hood. You just did not see its results.

Thanks for the info! I thought as much because some of the tracking issues I'm having are the same ones I had in 1.0. The errors occur when she crouches down very low to the ground and iPi Studio will completely loses the position of her legs--now I understand better why this happens.

To new users who are curious, this motion data is unusual because of the size of the performer and her quirky moves but it does present some interesting problems you may run into with adult performers. When I have time, I'll make a 'walkthrough' video to explain how different tracking problems were solved within iPi Studio. :)

G.

Greenlaw · **Posted:** Thu Aug 23, 2012 12:58 pm

After thinking about what you wrote regarding camera height, I had a thought:

In this example, the Kinect cameras were set at a normal height for capturing a full-sized adult performer but it just dawned on me that the cameras were probably positioned too high up for a small child performer. (She was almost exactly 1 m when this data was recorded, so approximately half adult size.)

Next time I'll try positioning the Kinects more appropriately for her height. The result will probably still present an issue with her 'thin-ness' but I think it should be better than what we see here.

Let you know how that goes.

G.

Michael Nikonov · **Joined:** Tue Mar 31, 2009 9:13 pm **Posts:** 805

vmaslov wrote:

In v1, essentially the same algorithm was used by the tracker under the hood. You just did not see its results.

Threshold value in v.1 for background was slightly different close to the floor, so theoretically there can be some difference between v.1 and v.2.

In v.2.0, threshold value for background still is different in the area close to the floor, depending on "background subtraction" compression method being enabled or disabled. Converting your old video to iPiVideo format with "background subtraction" disabled may produce slightly different results.

We will try to implement more accurate background subtraction in future versions.

When recording new videos with Kinect and iPiRecorder v.2.0, you may want to disable "background subtraction" compression method - so that you could take advantage of more accurate background subtraction in future versions of iPi Mocap Studio. (Provided that you actually are interested in re-processing your videos sometime in the future, and you have enough hard drive space).

Greenlaw · **Posted:** Thu Aug 23, 2012 2:21 pm

Thanks for the tip! I'll try that and compare results.

G.

Greenlaw · **Posted:** Fri Aug 24, 2012 12:53 am

I still had iPi DMC 1.0 installed so I took a fresh look at the original 'Three Happy Cats' data. You're right, the feet are clipped in that version too.

I will let you know how the new 2.0 data looks here after I get to recording it. At this point I'm ready to look at something new! (I still plan to do the 'walkthrough' video though.) :p

G.

Feet being clipped

Who is online