Samsung Files Patent for Faster Tracking of Real Objects in AR Apps
Samsung's patent application describes a way for AR software to work out how a real object is positioned and tilted without redrawing it over and over to check its guess.
What Samsung's object-tracking patent does
Today, when an AR app needs to know exactly how a real object sits in front of your camera, the usual method is slow: guess the position, redraw the object on screen, compare the drawing to the photo, and repeat. The filing says that redrawing eats computing power and adds lag, which can spoil the experience.
Samsung's approach skips the redrawing. The system picks out landmark points on an object, such as the corner of a table, and uses a small neural network to nudge each point to where it really is. From those corrected points it works out the object's position and tilt, then loops a few times to tighten the answer.
There is also a pre-step. Your camera photo is first restyled to look like the computer-made images the system was trained on, which helps it cope with real-world pictures.
… determining modified key point information based on a key point feature map corresponding to the key point information; and estimating an object posture of an object in the image, based on the modified key point information.
Translation: It refines spatial data using a feature map to figure out how an object is positioned.
How the key-point nudging loop works
The system starts with key points: distinctive spots on an object, like the corner of a table or the edge of a plate. A detector network (a neural network trained to spot things in images) reports where each point is and whether it is visible. Points hidden behind the object itself or something else are thrown out.
Next comes the correction step. For each remaining point, the system crops a small patch of the image's internal data (the filing gives 16 by 16 as an example) and feeds it to a refinement network. That network does offset regression, meaning it predicts a small x and y adjustment, and the system adds that adjustment to the original point.
The corrected 2D points are matched against a stored 3D model of the object, such as a CAD file or a scan. A standard math method called perspective-n-point then works out the object's six degrees of freedom: three numbers for where it sits and three for how it is rotated.
The result can be fed back in. The system projects the 3D model's points through the estimated position and the camera's built-in properties to update the key points, then corrects and estimates again. It stops when the points barely move between rounds or a preset number of rounds is reached. A separate claim covers the restyling step: converting the photo to the look of the synthetic training images before any points are found.
Why skipping the redraw step matters for AR
If you have ever watched a digital label or character slide off the thing it was supposed to stick to, you have seen the issue this filing addresses. AR only works when virtual content stays lined up with a real object as you or the object moves, and the filing says the slow redraw-and-compare step in older methods adds delay that hurts the experience.
The filing says the method could run on phones, tablets, game consoles, wearables and vehicle terminals, and that it can handle objects moving through a scene. It reports no speed numbers, so how much faster it would be is still an open question. The restyling step goes after a separate issue: models trained on computer-made images can stumble on real photos.
Samsung's 81st filing we've tracked in the AR glasses race since May follows work like the folding camera patent and the passthrough alignment fix.
The delay is real. An AR overlay that trails a moving object by even a moment looks wrong, and the filing points to the usual redraw-and-compare loop as a major source of that lag. Cutting that loop out goes straight at the cost.
The size of the fix is harder to judge, because the document gives no timing or accuracy results. The core claim is broad: find points, correct them, and estimate the object's position and tilt.
The photo restyling step goes after a smaller but familiar headache, since software trained on computer-generated pictures meets real photos that look different. It is a sensible answer to a real problem. The proof would be speed measurements that this filing does not offer.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Samsung patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
12 drawing sheets from US 2026/0350835 A1 · click any drawing to enlarge
Show the other 4 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in