Here are some notes I gathered studying and experimenting with the C A H V O R E model, using Gennery (2002) and Di and Li (2004).
Each letter in CAHVORE represents a parameter of the camera model which allows for projecting a 3d point onto an image plane. The model is based on a conceptual description of the projection and its parameters are tuned to a specific camera in a careful, experimental calibration process. It is important to understand that this process is a least square fitting minimization of residuals tuning all 7 parameters simultaneously. This means, while each parameters has a meaning in the concept to derive the model, that the result of the calibration is not strictly required to coincide with a physical measurement. In practice, I believe the parameters mostly do follow an actual measure because they use careful starting guesses and because there are some constraints on the calibration.
C A H V characterizes a perspective view with a focal distance and an orientation where straight lines stay straight. O is needed when the optical axis is a bit off the camera axis.
R stands for rho and is the radial lens distortion parameter. E stands for epsilon and is the fisheye parameter. E is needed because R cannot capture distortion of very wide angle (>140 or so) lenses.
The model first applies E by considering a variable lens pupil location which depends on the incidence angle of the entering ray rather than a fixed pin hole. This has the effect of an apparent forward shift of the camera center C resulting in an effective center C' and can thus characterize wide angle lenses. Note that the description of how E is defined differs in Di and Li (2004) and Gennery (2002). Di and Li (2004) only briefly mention E and do not use it. Gennery (2002) is the basis for the NASA/JPL use of the model.
Then R is applied by using it to determine an apparent shift in the position of the imaged 3d point P to an effective location P'. This shift describes the lens distortion. Finally, P' is projected onto the 2d image plane using the other parameters.
There are three types of the mathematical model: the general model, the fisheye model, and the perspective model. I believe they are distinguished for historical reasons and because they allow for simplified computations. The general model can represent the other models by using an additional parameter (L or P).
Since the calibrations occurs for all parameters simultaneously it is not really possible to apply the model partially. For example, for fisheye lenses both parameters R and E have to be applied. In fact, the exact mathematics described in Gennery (2002) should be probably used even if there may be other ways to describe the concept mathematically since it is those equations which are used for the calibration, at least these are the only ones which are accessible. Following these the model can be used to project a 3d point onto a 2d image, and Gennery (2002) describes a sanctioned process to do the reverse. However, only the orientation of the entering 3d ray can be recovered from a 2d pixel. Unfortunately, this involves numerically solving an equation for each pixel.
The MARSCAHV VICAR utility attempts to remove distortion by transforming an image from the CAHVORE model to the CAHV model. Apparently, the utility is not available publicly. From its description, it computes the orientation of the entering ray for each pixel as described, assumes a 3d sphere (of radius 1m) as the 3d point location, and then projects that 3d point onto an 2d image using the CAHV parameters.
I may try to reproduce that approach. If it works, an interesting experiment for the Nav helicopter images may be to assume a horizontal plane (normal to the camera axis) at 10m distance rather than 1m sphere to intersect the computed ray with. And then simply look at the map view, eg. take the x and y and ignore the z for an image.
Similarly, for the color RTE images, assuming a plane at a 22 degree angle to the camera axis for intersecting the reconstructed rays and looking at the map view could be tried.