Robotics Stack Exchange is a question and answer site for professional robotic engineers, hobbyists, researchers and students. It only takes a minute to sign up.
Sign up to join this community
I am currently considering to use monocular visual-inertial odometry (VIO) to localize a smartphone relative to a known object. Now, what I would like to understand is, what kind of accuracy I can expect with such methods.
Is there any rule-of-thumb, which tells me something like a standard deviation from the real position, assuming that the filter is properly initialized and we have average smartphone hardware?
There are certainly a lot of variables that would make the answer to this question "it depends" (e.g. lighting conditions, feature richness of the known object, generality of the algorithm being implemented, expected speed of motion around the object, compute performance, sensor quality...).
Smart phones typically have cameras with rolling shutters and low quality (noisy) MEMS IMUs. So if you are moving too fast with respect to the features that the VIO algorithm would track, you will likely experience serious drift, and the standard deviation of your prediction error is a moot point. The baseline for phone sensor types (mono, stereo, structured light), quality and compute power, as well as the algorithms being developed also keep changing, so its difficult to have a current, one-size fits all rule of thumb.
That said, I know there's still value in having some ballpark numbers. For that, my best recommendation is to skim through some recent literature to get some estimates for comparable constraints/situations.
No shortage of prior art out there, though admittedly many past works are focused on the robotics space where either: additional odometry sources (e.g. motor data) are available or more compute than is typically available on a mobile phone is used...
For instance, this 2022 paper compares Augmented Reality tracking performance of different devices while moving through a simple environment, the best of which achieve decimeter level RMS Relative Path Error and Absolute Trajectory Error.
This 2020 paper claims ~5cm accuracy for real time reconstruction room sized scenes.
This 2015 paper running on an iPhone 5 claims ~1cm tracking error while reconstructing nearby objects.
I don't want to dive too deep into any particular paper and won't claim that this is an exhaustive (or current) review of the state of the art, but hopefully that gives you the order of magnitude you were hoping for.
