Camera Drift in ARCore

[achuvm]

Hi,

We are currently using ARCore's NDK and both us and our users are noticing some drastic camera drifts without any notifications on ARCore's side of things.

What we see is that the camera position is being updated and moved farther and farther away even though the user is sitting still. During this time we are also monitoring the ArCamera's TrackingState and it is neither AR_TRACKING_STATE_PAUSED nor AR_TRACKING_STATE_STOPPED. As a result, there's no way for us, or our users to determine that the given positions are invalid.

I've attached the logs for one occurrence of this issue below.

If you take a look at the logs, you can see that the ArCamera starts off reporting Unavailable (our wrapper for PAUSED or STOPPED) and is returning [0,0,0] initially.

Once tracking initializes, the position seems to be stable for about 12 seconds before beginning to drift out to z=-45 all the while reporting that the camera is normal (our wrapper for ArCamera's Tracking state).

Once the z position hits -45, then the ArCamera begins again to report Unavailable before reverting back to normal, drifting out to z=-50, and then settling on z=-18.

The entire time this is happening, I am simply sitting down and looking forward at my monitor (there is a window to the side of it).

One peculiar thing to note is that the issue of drifting seems to happen to our Pixel and Pixel 2 (both are non-XL) devices a lot more than on our S8+ device we have here. With the S8+, this is rarely, if at all an issue, but with a Pixel, the drift seems to happen often after starting ARCore.

Logs for this issue: camera_drift.txt

[inio]

Could you capture a video showing this behavior, ideally with point cloud display enabled?

[achuvm]

Hi @inio,

This seems to happen more on app start and it is interesting that when it happens the cloud points are non-existent or "far" away. See 2 instances of this in the videos below.

Longer video, heavy shifting
https://drive.google.com/file/d/1gjZfgN5NX5LY4hmOGINAwX2pGDu4KZoh/view

Shorter video, more typical of what happens, slight drift and shift.
https://drive.google.com/open?id=1oBf-zUIpaBYN051lmzTQ5rzjgRyapPJu

In both cases, the white box is placed 1 meter in front of the user (in fixed world coordinates of [0, 0, -1], our camera starts off in the -z axis).

Thanks,

[inio]

Thanks for the videos! This drift definitely does seem kinda severe and I've sent to to the tracking team to make sure it's not indicative of something wrong, however:

If you want the lock something physically in place, you could create an anchor and not use world coordinates. Before creating an anchor, you should have enough of a map built to get at least one plane detection.

[dthian]

Hi @inio, thanks for looking into this. If it helps, I actually did a side by side comparison with:

1. A 3D droid OBJ model, placed in world coordinates (placed in relation to the AR Camera's transform, so that they move in reference to the camera's movement).
2. A semi-transparent black surface, representing Google's Trackable Plane (should react similarly to a Trackable Anchor).

Note that as seen in the video link i've attached below, we see that the both (1) and (2) are both drifting significantly, even though my camera was still and facing the same direction most of the time - You can see that both the droid and the plane "float/flys" almost out of view from the still camera.

As such these situations may affect Google's Trackables as well. And it's probably occurring because of the same issue that @achuvm is seeing above:

What we see is that the camera position is being updated and moved farther and farther away even though the user is sitting still.

ARCore Drifting Video Bug:
https://drive.google.com/open?id=1dM3DkZqX_IeWG5eO8JSG6BPx6GB0yLUu

[achuvm]

Hi @inio,

We are also seeing this issue using the Hello AR Java sample app provided here.

Here's a video for it:
https://drive.google.com/file/d/1Kxjr80NoR5Py962Orl1OxgrPqj0oj8SF/view?usp=sharing

Thanks,

[dthian]

Hi @inio,

Is there any update on this issue? Also, isn't this a bug, even though it doesn't the labeled "bug" like the other ARCore issues (should it have that label)?

The last video posted by @achuvm has a pretty dramatic drift and accurately demonstrates how this bug can drastically impact the user experience of customers using ARCore.

Thanks!

[inio]

@achuvm: that's pretty dramatic drift. Is that on the Pixel or the S8? Would it be possible to get a logcat dump from a session like that?

@dthian: Some of the early samples were less conclusive on "working as intended" vs "bug", but @achuvm's latest video is definitely bug-worthy behavior.

[inio]

@achuvm Ping. Any chance we could get a logcat of a session with drift like that? That seems like really bad calibration, but on a Pixel 2 especially that should be nearly impossible since each device is individually calibrated.

[achuvm]

Hey @inio,

Yeah, sorry, we've been busy. I'll try to get you a logcat by EOD.

Thanks for the ping!

[achuvm]

Hi @inio,

We were having issues repro-ing the issue today (the Sun is out in Seattle), but we did see a bit of drift at the end of the following logs:

drift.log

Observationally, tt looks like the issue occurs right before we lose tracking in that we'll see things start to drift, and if we don't point the device back on an easier-to-track surface, then the object will start flying away before possibly disappearing.

Thanks,

[inio]

@achuvm Thanks! I'll get that to our algorithms group.

[Edgar-BlipparDev]

The drift is sever when tracking is lost and person moves to different angle / location

[camilorodriguezapriorins]

Hi to everyone

We are experience this drift, too. This happens sometimes when you barely load the ARCore scene. Other times, it happens after a few seconds when the camera "freezes". Do you guys know if google has a plan to solve this in a near future? @achuvm or someone, have you find some workaround to minimize this drift?

[peetonn]

Stumbled upon this thread because I might experience similar issue (described on Unity Forum). In my case - point cloud is drifting, not sure why. Here's the video.