Accuracy of GPS vs VPS

You might have noticed that GPS accuracy is commonly measured in meters or centimeters because this is the most useful way to express the physical distance between the true location of an object and the location determined by the GPS satellite.

So, in consumer applications like in-car navigation systems or fitness trackers, the GPS localisation accuracy is typically around 5-10 meters. What that means is that the satellite estimates your location to be within a roughly 5-10 meter radius. And this is sufficient to provide reliable turn-by-turn directions, identify roads, calculate distances, and guide users effectively.

Similarly, the accuracy of VPS (Visual Positioning Systems) localisation today is also expressed using centimeters. 

So when an Augmented Reality experience is anchored to a specific location in the physical world, people often say, “This is accurate to within 2-3 cm.”

So what do we mean by that?

Well, in most cases, we are talking about the spatial map (a digital representation of the surfaces and shapes in the physical space). We are basically estimating how closely the points or mesh structure of the spatial map align with the physical objects.

This can be useful when we are overlaying digital information on top of a museum exhibit, or a plumbing pipeline, because the viewer is expected to be standing a meter or so away from the physical object. So when we say, the digital content aligns within 2-3 cm of the physical objects, that might be true because each pixel in the image (viewed through the camera) represents a very small distance in the physical world.

But if we are anchoring digital content to Tokyo Tower, or T-Mobile Park baseball stadium, the viewer would be much further away — and although the AR objects might still be aligned to within “a few centimeters” from the user’s point of view, we cannot say this is strictly true because a single pixel could represent 20 meters of physical distance.

I think there is a better way to measure the accuracy of VPS localisation. 

The key component in VPS that does not exist in GPS is the camera. Or, I should say cameras because there are two variations.

Physical device camera: the actual camera hardware on a physical device, such as a smartphone, tablet, or specialised camera equipment.

Virtual software camera: the simulated camera within a software application.

With that in mind, I would like to propose that a more effective way to measure the accuracy of VPS localisation is by examining the angular accuracy — how closely the orientation of the virtual camera matches the orientation of the physical camera of the device.

If the device camera is pointing in a certain direction in the physical world, the virtual camera should ideally point in the exact same direction in the virtual space.

When you move your phone to look at an AR object, the app should accurately render the virtual objects in alignment with the real-world view seen through the phone's camera. If the virtual objects are slightly misaligned (e.g., appearing slightly to the left or right compared to their intended position), this misalignment can be measured in degrees.

If the physical camera is pointing exactly north and the virtual camera is pointing slightly northeast, there might be a 5-degree difference. This 5-degree difference represents the angular accuracy. The smaller the difference, the more precisely the virtual camera mirrors the physical camera's orientation.

The human perception of AR content relies heavily on the correct perspective. That’s why I think angular accuracy is a more useful way to measure the integrity of an immersive experience. It scales better across different environments and remains effective whether the content is being viewed up close or from a distance.

If you want to chat more about low cost localisation or spatial computing, please do not hesitate to reach out to me.