[BugsJustFindMe]

I have a degree in Computer Vision, and, whether Pickle is lying about various capabilities or not, this guy is talking completely out of his ass and a whole lot of what he says is just extremely idiotic.

> tracking blah blah 6DoF blah blah IMU

This whole section is just wildly false. Tracking like shown in the video is easily done with just a camera, 1980s-era sparse optical flow, and basic fucking geometry. No IMU needed. People have been doing far more complex and stable motion tracking with no more input than single camera video for literally decades. And this device doesn't just have a camera; it has two HD stereoptic cameras, so they also get a depth map. You can absolutely do what they show with the hardware that Pickle claims is in the glasses.

(If you want a fantastic example, see the intro sequence to the movie Stranger Than Fiction from 2006.)

> It would take time to affix an open source SLAM pipeline and even more for them to build their own.

And this is a complete non sequitur, as SLAM is also not needed for what they show in the video. Nothing shown requires mapping the area. It's also a joke to say that it would "take time to affix an open source SLAM pipeline" unless by "time" he means a few minutes.

> This would indicate either the software is using real-time depth tracking blah blah

The glasses have fucking binocular cameras in them! What the fuck else would they be for?

> But in the photos of Pickle 1, there is no sign of any spot to charge the device.

There is zero reason whatsoever to believe that those images are photos of the final product and not renders or props. It's like he's never seen marketing material before.

I can't even with this.

This guy's LinkedIn bio says "Aug 2022 - Mar 2023: Attended UVA as a first year studying economics and commerce before dropping out to build in VR full time." So it seems he's a self-important child with zero background. That explains a lot tbh.

[charcircuit]

SLAM is needed for world locked content.

>1980s-era corner feature detection, and basic fucking geometry

Which are pieces of how SLAM works.

>You can absolutely do what they show with the hardware that Pickle claims is in the glasses.

World locked content is not novel. Existing glasses can do it today. The claim is that Pickle didn't build it. The obvious answer would be that they are using what Qualcomm or someone else built it as opposed to Pickle building all of this within a month.

[BugsJustFindMe]

> SLAM is needed for world locked content.

It absolutely is not. Tracking is needed for mapping, not the other way around.

And it's definitely not needed for what they show in the video that this kid is complaining about. It's not even needed for associating things that go out of view and then come back, though it can help there.

> Which are pieces of how SLAM works.

Screws are pieces of how automobiles work, but it would be foolish to suggest that one needs a Honda Pilot to hang a painting on their wall.

> The claim is that Pickle didn't build it. The obvious answer would be that they are using what Qualcomm or someone else built

Please don't shift goalposts. The claim is that they're lying about capability. And the evidence given for that claim is flat out wrong.

[charcircuit]

>Tracking is needed for mapping, not the other way around.

The term tracking is being used twice. The tracking data that OpenXR exposes comes from SLAM. SLAM is done doing sensor fusion including signals that come from tracking points.

>And it's definitely not needed for what they show in the video

The video shows 6 dof tracking which for a production implementation would do SLAM tracking.

>for associating things that go out of view and then come back

Having memory of what existed before implies you have a form of a map. You also want a map to be able to match together the views of the multiple cameras.

>Please don't shift goalposts.

The claim I am referring to is, "6DoF with spatial anchoring on a device this small and compute constrained is hard for any company to build, let alone Pickle."

[BugsJustFindMe]

> The video shows 6 dof tracking which for a production implementation would do SLAM tracking.

SLAM is not required to do what is shown in the video. As is an IMU. And an IMU is also not required for SLAM. Everything about the blog post is factually wrong.

> Having memory of what existed before implies you have a form of a map

Once again, you're just wrong here. Image feature correspondence works even without any spatial mapping. Once again, you're getting things backwards. You need to find correspondences before you can begin to make a map, not the other way around!

Anyway, I don't have the energy to argue more with someone who confidently doesn't actually know what they're talking about. So, good luck, have fun.

[brcmthrowaway]

> This whole section is just wildly false. Tracking like shown in the video is easily done with just a camera, 1980s-era sparse optical flow, and basic fucking geometry. No IMU needed. People have been doing far more complex and stable motion tracking with no more input than single camera video for literally decades.

Not with imperceptible latency

[BugsJustFindMe]

(nb. I probably should have said sparse feature tracking and not optical flow. People tend to get the wrong idea about what optical flow fundamentally requires. Spatial regularity and density are not inherent but people may assume they need to be.)

First of all, did you watch the video? (the whole thing is kinda annoying and long, but the part in question here is only about 3 seconds so it's worth looking) Two points about the video: 1) The positioning of the overlay is noticeably unstable in relation to the apparent camera motion, so it doesn't even show what the OP claims it does. 2) You don't have any way to know what the latency is because of that.

Anyway, yes even with, and even in the form factor if you optimize for the right things. The kind of simple feature tracking that can accomplish what's shown in the video was real-time in like 2005, and there have been significant hardware and algorithmic advancements in the past 20 years.

[NBJack]

Given the amount of tech I own that is supposed to do this (higher end VR headsets with hand tracking, AR glasses with environmental tracking, etc.), I wouldn't dismiss the author's claims.

But I'd be interested in your examples that can achieve what Pickle is offering in a single pair of glasses.

[fuzzythinker]

If you really believe Pickle's claim so much and disbelieve OP's analysis so much, you should contact Pickle's CEO to get more info from him. Once you build more trust, you should join the CEO and take on OP's bet.

[BugsJustFindMe]

Pickle's capability claims don't even need to be true for OP's "analysis" to still be extremely factually incorrect on many levels. Also I consider gambling to be on the balance a bad thing and have no desire to encourage it.

[BugsJustFindMe]

(I probably should have said sparse feature tracking and not optical flow. People tend to get the wrong idea about what optical flow fundamentally requires. Spatial regularity and density are not inherent but people may assume they need to be.)

[ruufu]

Well I hope you never get into this business because I doubt your glasses will work in the dark or on red eye flights.

[BugsJustFindMe]

I have a sneaking belief that the video clip showing a rack of cookies on a brightly lit counter wasn't shot in the dark or on a red eye flight, so that's a non-sequitur.

Digital camera sensors are all inherently extremely sensitive to infrared anyway and can see quite well in the dark with nothing more than an IR LED if you don't add a physical filter over the sensor, soooo...