Camera Models

cameras.hpp + cameras/* provide the projection models the front end uses to map between 3D camera-frame points and image pixels. Each is header-only, type-generic, and provides project / unproject / distort / undistort plus the analytical d(pixel)/d(point) Jacobian.

The three models

Model	Header	Use
Pinhole + radial-tangential	cameras/pinhole_radtan.hpp	The standard EuRoC MAV model (Brown–Conrady 5-parameter distortion). The default for the VIO benchmarks.
Equidistant fisheye	cameras/equidistant.hpp	Kannala–Brandt model for wide-FOV lenses.
Unified omnidirectional	cameras/unified.hpp	Mei–Rives model for catadioptric / very-wide cameras.

Pinhole + radtan in detail

PinholeRadtanCamera<T> carries intrinsics (fx, fy, cx, cy) and a 5-parameter radtan distortion (k1, k2, p1, p2, k3):

• project(p) — normalizes (x/z, y/z), applies distortion, scales by the intrinsics → pixel.
• unproject(px) — inverts intrinsics, undistorts, and returns a unit bearing (camera frame, +Z forward).
• project_jacobian(p) — the analytical 2×3 d(pixel)/d(point).

The default-constructed camera is the identity pinhole (fx=fy=1, cx=cy=0, no distortion), for which unproject round-trips normalized coordinates. This is what lets tests and pre-rectified front ends feed normalized coordinates directly while production code supplies a real calibration.

How the estimator uses them

The MSCKF core never sees these models directly — it works in normalized image coordinates (see Layering Invariants). The VioEstimator holds one camera model per camera id and uses unproject to convert incoming pixel observations to bearings before handing them to the camera updater. Swapping a pinhole for a fisheye is therefore a front-end configuration change, invisible to the estimator math.