Forum Discussion
Establishing ground truth / validating sensor fusion
Hey folks,
I want to start experimenting with mobile device sensors and sensor fusion to estimate pose. This raises the question of how to do validation. Years ago, I tried just double-integrating iOS linear accelerometer data and concluded that my naive approach + the hardware wasn't good enough at that time.
So, coming back to it, how can I rigorously establish ground truth to compare my estimates to? I've seen pictures of a robot arm at Oculus -- but that's a bit out of my price range. Anybody have experience with this?
I'd be willing to start with movement along one axis and/or rotation about one axis only. Maybe I could get by with a custom-machined jig that has a marker at a known distance to some known tolerance?
Thanks in advance for any thoughts and tips!
I want to start experimenting with mobile device sensors and sensor fusion to estimate pose. This raises the question of how to do validation. Years ago, I tried just double-integrating iOS linear accelerometer data and concluded that my naive approach + the hardware wasn't good enough at that time.
So, coming back to it, how can I rigorously establish ground truth to compare my estimates to? I've seen pictures of a robot arm at Oculus -- but that's a bit out of my price range. Anybody have experience with this?
I'd be willing to start with movement along one axis and/or rotation about one axis only. Maybe I could get by with a custom-machined jig that has a marker at a known distance to some known tolerance?
Thanks in advance for any thoughts and tips!
4 Replies
- Yes, the guys here have used robot arms but clearly that could be... er... "out of reach" to hobbyists.
You could probably do something with AR markers stuck to the phone. Not sure if that's enough precision though. - I'd be willing to wager that for the price, the DK2 camera tracker is the most precise ready-made 3D pose tracker that you can get. I would tape the phone right onto a DK2 and gather ground truth like that.
"brantlew" wrote:
I'd be willing to wager that for the price, the DK2 camera tracker is the most precise ready-made 3D pose tracker that you can get. I would tape the phone right onto a DK2 and gather ground truth like that.
The problem with this is that I don't have any data on the accuracy of the tracker. Even if it's the best tracker, the whole point is that I want to quantify the error.
Maybe I can get a surplus rail & slide from a machine tool to get good data on one linear axis. Alternately, I could build a lego rail & gear system and use that with a metal straight-edge. At first I was thinking that I'd need components that had been ground flat, but if I'm manually cranking the slide and reading the position off a ruler or measuring it with calipers, then I can probably get < 0.5 mm with a cheapo setup. It doesn't have to be flat, because I only care about motion along one axis.
EDIT> Huh, actually I'm coming around to this idea. Thanks for the suggestion. If anything, I can clamp the DK2 to my cheapo rail and figure out whether it has the accuracy I want. Move the DK2 0.5 mm and see what it reports...- I use a turntable for Gyro calibration with a paper map of compas points, and degrees around a circle(http://www.amazon.com/gp/product/B0045DV04I/ref=oh_aui_detailpage_o03_s00?ie=UTF8&psc=1). This works well with my IMU, I am sure you can do something similar with the rift, but I have just taken it for granted it was calibrated .... :oops:
~B
