Hi IceDreamer,
Thanks for sharing your benchmark results. Very interesting.
Older posts in this forum have mentioned that a key point limit larger than 40,000 and a tie point limit of 4,000 does not add much to the overall accuracy. Your results seem to match these observations (curve flattens at ~4,000 tie points). Note that these optimum key point and tie point values might change if you use photographs with very high resolutions (larger than 36 Megapixel). What was the resolution of your images?
Regarding the accuracy settings. If you choose HIGH, Photoscan is using all images at FULL RESOLUTION, when you choose MEDIUM it is using photos at 50% and when choose HIGHEST it is actually 'upscaling' your images to 200%. That's why I generally don't use HIGHEST for Photo Alignment because I am not really sure what this 'upscaling' process does to the project (as somehow confirmed by your results). HIGHEST might be helpful for very low resolution (less than 2 Megapixel) imagery, where otherwise SIFT wouldn't be able to identify enough key points (some multispectral cameras have quite low resolutions, for example).
Therefore, it might be a quite difficult to properly 'interpret' your orange line.
Ignoring the orange line, let's have a look again at your graphs.
Mean key point size for HIGH is lower than for MEDIUM. Makes sense.
RMS projection error seems to be better (lower value) for MEDIUM than for HIGH, but if you consider that we are already talking about a sub-pixel error (~0.20 vs ~0.35 for 4,000 tie points) for images where the smallest unit is 1 pixel, then I think both of them are very good.
Another check if your overall accuracy between HIGH and MEDIUM is changing significantly would be to use (properly surveyed) markers (check points & control points). Assess their errors.
In case you haven't already done so, run OPTIMISATION after Photo Alignment and after removing points that show a a reprojection error higher than 1 (EDIT > GRADUAL SELECTION). I think this will further improve your accuracy.
Regards,
SAV