Here's what actually issued. On July 6, 2021, Baidu USA LLC was granted US11055540B2, "Method for determining anchor boxes for training neural network object detection models for autonomous driving," inventors including Ka Wai Tsoi and Tae Eun Choe. The CPC codes are vision-recognition classes (G06K 9/00791 for road-scene recognition, plus G06K 9/6202/6218/6256 for matching and training).
The mechanism is about anchor boxes — a workhorse trick in object detection. Detectors don't search every possible box in an image; they start from a set of predefined reference boxes (anchors) of various sizes and aspect ratios, then refine them to fit real objects. Choosing good anchors matters: anchors that match the typical shapes of cars, pedestrians, and signs make the detector train faster and detect better. The claim covers a method for determining those anchors for the driving domain specifically.
Why patent the anchor-selection step? Because in autonomous driving the object distribution is distinctive — vehicles and pedestrians at characteristic scales and viewpoints — and a method that tailors anchors to that distribution is a concrete, defensible engineering contribution. Baidu, which has run a substantial self-driving program, has clear reason to protect the perception-training pipeline.
On scope, the usual line: granted B2, enforceable, but the claims cover this specific anchor-box determination method for the driving use case. They do not reach object detection broadly, nor anchor boxes as a general concept, which long predate this filing. Read claim 1 for the actual boundary.
The takeaway: US11055540B2 is a clean example of domain-specific AI IP — not a new architecture, but a method that adapts a standard detection technique to the demands of one high-stakes application, owned by a company that operates in it.