There are two issues with the default point2dem grid:

1. The current algorithm sets the grid size as the "typical" ground sample distance, that is, the estimated value of how far two camera pixels are when projected onto the ground. That is not right, a good rule of thumb is that the DEM resolution is about 4x this ground sample distance.

2. Currently, this "typical" ground sample distance is obtained by taking the median of all estimated distances between two neighboring pixels on the ground. That usually works, but for LRO NAC and other linescan cameras, the ground sample distance in one direction and the other can be very different, e.g., 0.5 m along a scanline, and 2 m across scanlines. The median of a set of numbers some of which are 2 and some are 0.5 will be one of the two, while the correct solution here is more like 1 m. I think given the list of all GSD values encountered by poinit2dem, one should take the ones in the 25% to 75% range, and average them, rather than taking the median. This will avoid worst outliers and the final result in this example should be on the order of 1 m.

There is already some kind of logic like this for mapproject, maybe it can be factored out and used in both places.