Hi

I've just started using iPiSoft, and there's a problem with the calibration that I don't understand. I have three PlayStation3 cameras, and before spending more money to this, I would need to prove that I can actually make this work. I tried to set up the actual cameras just like in the istructions here. It seems, however, impossible to set the virtual cameras right. Whenever I move one camera, others follow (and thus go to wrong places). I have no idea, how important it is to get all camera positions exactly right. Where is the origin, by the way? I could try to put the camera positions manually in the xml-file, if necessary.

I managed to get two of the three virtual cameras set up correctly and run the calibration phase with "autodetect initial camera postions", but after the calibration was finished, the software claimed that distance between #1 and #2 cameras was 1.000 m (actually 2.5 m). After correcting that, it moved the virtual cameras too high. Therefore, the calibration is useless and I cannot make any progress. I have, of course tried, but "refit pose" never works since I cannot move the skeleton into a correct position (i.e., near the actor in the video).

With Kinect, everything worked just fine (since no calibration was needed). It just seems too limited for motion capture right now...

You're doing it the hard way. :)

You should not have to move the virtual cameras at all, especially when using PS3 Eye cameras. IMO, moving the cameras manually almost guarantees failure.

Automatic calibration with PS3 Eye is actually much easier and more reliable. All you need to do is a record good calibration video and write down the height of at least one camera. You can see an example of a calibration shoot here:

Calibration Video

Next, import the video and simply click Calibrate. To set the scale of the scene, enter the actual height of any camera (usually that would be camera #1). Save the data and load it when you're ready to track motion. And that's it.

Successful calibration depends largely on how well you set up the calibration video. If you take the time to do this right, the whole process shouldn't take longer than 10 minutes and the same calibration data is valid for every motion video shot using the same camera setup. There are more tips about this elsewhere in the forum.

Hope this helps.

G.

_________________
Greenlaw
Artist/Partner - Little Green Dog | My Demo Reels (2013,) (2015,) (2017,) and (2019)

Watch a one minute excerpt on Vimeo now!

Thanks for the tip. It is a little bit silly that you can alter the camera positions if it leads to a certain failure. I added a fourth PS3 Eye (the local GameStop must be amazed how well second-hand PS3 cameras are selling!!!) but at least with half circle adjustment it didn't seem to ease the tracking problems. I do not have enough USB cables to go ahead and test full circle though, and maybe not enough space either. The laptop I'm using is not going to be able to handle any more cameras when recording (can get 50FPS using 4 cameras) due to the lack of USB root hubs. So, I'm stuck with this setup now. The problem is, that I need to prove that this is actually something worthwhile spending money for and while I'm personally quite amazed what I have accomplished in such a short time, I'm afraid that it is not going to be enough.

The main problem seems to be that tracking feet/hands fails almost certainly every time you turn around or when you are in such position that one of the cameras cannot see where your limbs are. A dislocated shoulder syndrome seems to be visible in 50% of the videos and the rest have crossed legs syndrome...

Here's something I've thought I should try, any idea what would be the best solution to try?

Lifting of the cameras as high as possible (abt. 8 ft/2.5m above the floor)
Using full circle instead of half circle positioning
Trying to level the cameras (horizontally) more accurately
Trying to wear trousers with totally different color pant legs (not sure if the software will even accept this)

Lifting one camera up is the easiest thing to do, so I'll start with that since it doesn't require more than putting the tripod on the table.

P.S. This forum logged me off while I was writing this, luckily I could save the text using "Back" on Firefox. Nevertheless, very annoying. Should I not log in from two different browser tabs or what was going on?

I think the software allows you to move cameras so you can give it general hints about which cameras goes where, just in case the system is getting mixed up about camera order for some reason. This hasn't happened to me in about two years though.

I'll try to comment on the rest of your post later. Not sure why your browser is having problems; this website is just a standard PhpBB forum. It will log me out after a long period of inactivity but certainly not within a few minutes.

G.

_________________
Greenlaw
Artist/Partner - Little Green Dog | My Demo Reels (2013,) (2015,) (2017,) and (2019)

Watch a one minute excerpt on Vimeo now!

Well, it seems now that the biggest improvement for tracking was to set the actor to -10-15 cm below the actual height of the person. Another result produced by shortening the actor was that previously the back was always bent a little bit and now it is straight. Even though this means that during T-pose the actor's feet are now too short (about 5 cm above ground) the tracking of the feet seems to work as before. I have no idea why, maybe my calibration is still not quite right? In addition, setting the shoe color to 0,0,0 instead of some gray level (which was produced by analyzing the actor) helped tremendously tracking the feet correctly, this was before making the actor shorter, though.

One extra wish that I have is that the full version of the iPiStudio could have the capability to process several files without attention from the user. In practise, this would allow capturing session during the day, and processing during the night when there is no other use for the computer. I would also be interested about the possibility to use SLI (two GPU:s in parallel).

That seems very unusual. It's generally recommended that the iPi rig be slightly larger than the performer (a few cm,) and the feet should certainly be on the ground, not below it. I think there's more info about this in the Wiki.



What you're asking for is a batch queue. I've asked for this too, and yes it would be a nice convenience. Hopefully it's coming one of these days but personally I would rather see head tracking implemented before that. :)

_________________
Greenlaw
Artist/Partner - Little Green Dog | My Demo Reels (2013,) (2015,) (2017,) and (2019)

Watch a one minute excerpt on Vimeo now!

I wonder, if a batch queue is not possible at the moment, would it hurt to run two instances of iPiStudio simultaneously on the same computer? At least that way one could use some of the idle computer time out of office hours... We did some recording tests outside of our office today in order to see whether it is the limited space we have at the office that is the cause of all our troubles. However, now there was some weird problems with iPiRecord, too. One of the cameras froze constantly in the middle of recording and that piece of recording was almost always wasted. The extension cable for that camera is brand new but it didn't give any problems before at the office. Second problem which seemed unrelated was the I/O-buffer that kept running out during the afternoon takes. This hasn't been a problem before. It went away when the laptop was rebooted but soon it came back. Maybe the hard drive gets slower when it is near full (like 80-90%)?

Just a couple of today's captures have been processed but it seems still that the main problem persists: the arms (mainly shoulders) and legs get very easily twisted to the limit and cannot follow the movements of the limbs beyond that limit and end up in very funny positions. I've always kept "shoulder tracking" on since I don't understand what would be the situation not to do so.

At least we would much more happily wear some markers during the capture than trying to guess what color clothes we should wear. The Actor -tab has some odd features as well, it doesn't always remember that actor was wearing t-shirt over long sleeve one. And "analyze actor appearance" gives quite astonishing results also. We are pale-faced but the analysis results to a very dark skin for the actor. And Hands = Face color doesn't seem to do anything (at least nothing I can see). I've tried to manually set the colors more correctly for the clothes but it is frustrating when you don't have a clue if it is better of worse (randomly it seems to matter, not always though).

iPi actually recommended that a while back. Of course you can only do that on one computer so you will be limited by RAM and processor speed.



Is this a USB repeater cable (aka, active USB cable)? These are different from standard USB cables in that they have a unit at the end that boosts the data signal. You may already know this but it's highly recommended that you use a repeater cable when extending the cable for a PS3 Eye camera. (And probably for the Kinect if you need to extend the cable.)

[url]Maybe the hard drive gets slower when it is near full (like 80-90%)?[/url]

Are you recording to the system drive? Unless it's an SSD, this is generally not recommended. I record to a separate SSD in my main computer with PS3 Eye cameras, and when using a laptop with Kinect I record to a USB 3.0 drive through a USB Expresscard.



Just a guess but I don't think markers would help because iPi DMC is apparently creating a volume based on the whole body. The use of markers may require a entirely different tracking system.


I've seen it sometimes forget longsleeve/shortsleeve settings, and yeah it's a little annoying. As for the iPi selected colors, this is really just to help iPi's ability to track the body parts as it sees them in the video, so you can't be too critical for it's 'tastes'. :)

If you run into problem areas, what may help is to alternate between analyze actor and refit pose a few times as you proceed through the timeline. This may be necessary if the lighting isn't well controlled in the room or if lighting changes dramatically as the performer moves near and away from light sources.

Ideally, you should try to light your shoot space as evenly as possible, but of course this isn't always practical or even possible. That said, iPi DMC is normally pretty good at dealing with small lighting issues if the video is otherwise shot well. The key is to getting good predictable tracking results is to minimize potential problems as much as possible.

Could you post a screen grab of a 'problem areas' in your video? It might help us here understand what you're dealing with.

G.

_________________
Greenlaw
Artist/Partner - Little Green Dog | My Demo Reels (2013,) (2015,) (2017,) and (2019)

Watch a one minute excerpt on Vimeo now!

Thanks for the answers!

We do have active USB cable extensions and yes, we do have to record to the (non-SSD) system disk. It is peculiar that the i/o problem started so suddenly, without any warning. Still, I do believe you... we need some alternative solution for that.

Good. We happen to have a desktop PC that has Intel i7, Geforce 560 with 2GB of memory. Our only concern was, how well multiple instances of iPiSoft will behave using the resources of one GPU (I mean, it wouldn't make much sense to track two videos in parallel, if the performance of both processes drops dramatically). Needs to be tested, then.

Well, I'm only concerned about the colors because there is the possibility for the user to adjust the RGB values. If after analysis the actor has a brown shirt (even though it is red in real world) should I just leave it that way or try to make it more realistic. All colors are usually brown, gray or almost black after the analysis except hands which always have pinkish color.

I uploaded a video of a shoulder twisting problem. The source of this problem might be the imported target character since at least I cannot see any problem until the skeleton is rigged. One odd thing I noticed is that the limits between iPiSoft default skeleton and imported character doesn't always match. You can see it here yourself. Maybe that is also the reason why certain parts of the skeleton will sometimes end up in funny positions?

Here's another example of an elbow which has bent way beyond the allowed maximum limit. The main problem with this is that now IK doesn't allow the user to move the elbow at all now, so manual corrections to the hand are very hard to do. Sometimes you can get the locked joint released by selecting rotate and fiddling with the spheres long enough but not always.