Calibrate Camera (Depth)

How it fits

Typical backends

Caveats

• Hard constraintThe vanishing-point path requires a Manhattan scene — visible families of parallel and mutually orthogonal straight lines such as building edges, floor tiles, doorways or road markings. A scene with no straight-line structure yields no vanishing point, so the estimate holds rather than updates.
• Hard constraintWhen the optional detection tail is wired, labels and sizes must be the same length — each entry of sizes is the real-world height and width in metres of the class at the same position in labels — and a mismatched length fails startup.
• AccuracyThe model is a square-pixel pinhole, so the recovered fx and fy are equal and skew is always 0. The emitted CameraIntrinsics carries the estimated radial distortion natively in its distortion field as the division-model [k1, k2] vector (k2 is 0); a scene with no measurable bow reports [0, 0].
• State lifetimeThe intrinsics estimate is a running value carried across frames, blended by decay so recent frames dominate; it converges over a stream and resets only on restart.
• I/O contractThe variadic input tail accepts zero, one, or both of a [BoundingBox] detection stream and a MetricDepthMap; absent tail inputs simply disable known-size fusion and never degrade the vanishing-point estimate, and every frame emits exactly one output.