Leap Motion - Is it supposed to be this jumpy?

We released a new version of VR Intro late last night that resolved some reported crashing issues. Can you download and try the very latest build?

"alexcolgan" wrote:
We released a new version of VR Intro late last night that resolved some reported crashing issues. Can you download and try the very latest build?

The new build works without crashing. And it supports Direct Mode. Yeah!

Finger tracking still loses alignment with my real fingers more frequently than I wish, but hand tracking mostly works. The stars demo was really really cool. I still worry that the tracking uses too much CPU to have enough power left over to use the Leap for game controls. Further, I really need finer motor controls for my current project.

Also, when you build out the Dragonfly, is there any chance that you will offload conversion of imagery into bone/point data? It would be awesome if the CPU did not have to perform this in software.

Here you can see some precise movement with the latest version.
Ensure that youve calibrated it properly and that you've got a good CPU and GPU for good results and low latency.




Cheers
Peter

"thorohtech" wrote:
Here you can see some precise movement with the latest version.
Ensure that youve calibrated it properly and that you've got a good CPU and GPU for good results and low latency.
Cheers
Peter

The video you presented has a lot of hand jumpiness as well.

This video shows the issue I am talking about:
https://www.youtube.com/watch?v=hxGTeNlsdcE


At about 4:00 you will see them trying to precisely drop a piece and it not working. You will also see hand jumpiness as well.

But, this video (from the developer) shows the opposite:
https://www.youtube.com/watch?v=1liHN32nUMo



This makes me wonder if some Leap Motions are defective and others work great.

Lighting really makes a HUGE difference, while there does seem to be some 'automatic' lighting calibration but I just use mine in the dark and I get some really good precise movement 'most' of the time. The jumpiness I get I think seems to be from the Leap not being able to tell if my hand is palm up or not, so I am looking for a way to intercept that with data from an IMU...well see I guess, I dont know if it possible or not, although I was having 'some' luck with IR dots, so Im ight try that again.

My experience with the leap has been disappointing. It's good enough to be tantalizing, but not good enough to use for anything. My biggest gripe is that fact that when it fails, it fails unpredictably.

One thing I've been meaning to try is to post-process the hand configurations that it returns. It seems that you could at least reject obviously bad configurations. I.e. "broken fingers", although this would add latency. I'm thinking like joint limits. If the current configuration violates the joint limits, then just reject it and keep the previous good estimated configuration.

Maybe also velocity limits. I.e. I'm a human, not an ant / spider. My hand doesn't rapidly oscillate between wildly different configurations.

I'm going to try the calibration tips, here, though. Will report back on whether they helped. Basically, I've never been able to achieve the reliability shown in leap's video demos, but I suspect that these are cherry-picked.

EDIT> I know this is a hard problem, but it's funny they seem to have solved so many problems -- they got into retail stores, the industrial design is nice, the box-opening experience is nice ... so many things done right from a product / industrial design point of view but the hardware + software integration seems little better than a research prototype.

Considering the amount of CPU that is eating up to see hands, I am not sure how much more post processing we can do.

Following are screenshots from the visualizer and the task manager showing CPU utilization side-by-side with latency information:

No hands:
https://drive.google.com/file/d/0ByrtvXdsXpmcWEJhdEJiNnFFemM/view?usp=sharing

One hand:
https://drive.google.com/file/d/0ByrtvXdsXpmcakRFc1pJQ2ljMXc/view?usp=sharing

Two hands:
https://drive.google.com/file/d/0ByrtvXdsXpmcd01aNTlsV21vSEU/view?usp=sharing

I am currently hoping for some feedback from alexcolgan as to whether this indicates faulty hardware or something else. When I see videos of people using the device perfectly, I certainly get my hopes up that it will work that well for my project :D.

Edit: The biggest concern I have is that some users have a great predictable experience while others have a bad experience. How do you code for that?

Looking closer at your earlier post, I failed to realize that you were describing the combined CPU usage of the Leap Service and the Visualizer -- sorry about that. For the Leap Service, 9% CPU with 1 hand and 16% CPU with 2 is fairly normal for a mid-range laptop.

The Visualizer is strictly a diagnostic tool, and its CPU usage isn't indicative of the resources needed to run the controller. The only CPU usage that matters is the Leap Service, which is what needs to run during a game. As long as a computer fits our recommended specs, it should be able to run graphics-intensive VR experiences with the Leap and the Rift.

With that being the case, it sounds like the calibration is the biggest issue. When you were calibrating, did you rotate the controller around and keep it generally within the same position, as shown in this video?

"alexcolgan" wrote:
With that being the case, it sounds like the calibration is the biggest issue. When you were calibrating, did you rotate the controller around and keep it generally within the same position, as shown in this video?

That is exactly what I did. But, it seemed like it took forever to eventually reach 80%. Certainly more than 2 minutes. Further, it kept dropping back to zero and I had to retry several times.

In the Leap Motion Control Panel, can you submit a software issue report through the troubleshooting tab? This will send along your device's tracking data and help us determine whether the hardware might be at fault. Be sure to hold your hands above the device while recording, and include "andrewtek" in the report so we know it's you :D