The official PixInsight reply is 'it depends'. But let's see on what it depends.
The ImageIntegration is very well document, even if the documentation needs multiple reading. You may just glimpse at the Figure 4 — Sigma Clipping Rejection and the surrounding text for a first idea of how the sigma works.
Rephrased in an approximate way, the sigma parameter is how far the value of a pixel in an image is of the median value of the same pixel on all images. All pixels further away than that value are assumed to be erroneous pixels and ignored (for example satellite trace). The reality is somewhat more complicated, but this give the idea.
The larger the sigma parameter value, the more pixels values are accepted. You want to accept as many values as possible (this is why you took the images in the first place), but not the bad traces.
It is called sigma because it is the name of the statistical property used to measure the distance. You do not have to reason in ADU or pixel values, you just have to know that a value of 3.00 should accept most of the values, but not all. Increase this value of you do not see any trace in the integrated image. Decrease it if you must remove more data (alternatively switch to another rejection algorithm). If you have a lot of very good image (like) you can really push the value pretty high (up to the limit). But more realistic values are between 2 and 4.
A very practical way to check if your result is clean is to first do an integration with the median (there is no rejection parameter as it rejects as much as it can, the image is noiser). Then you can superpose the median image and the integrated image and blink them with Ctrl/Next. Any non rejected trace will be obvious.
You can also look at the log after the ImageIntegration. There will be a line with the percentage pixel rejected. For an image it should probably be less than 0.1% (even a lot less for very good images), but the real check is the image, not the statistics.
So, indeed, it depends :-)
-- bitli
