The trouble with curved orders is that when the object projects to only one or two pixels on the detector, the bulk of the signal will sometimes fall into one pixel and sometimes it will be split between several pixels. This means that there is no single correct number of rows of data to extract, forcing the extraction of unwanted sky as well as wanted signal.
Obviously it would be possible to provide a program that would make a sensible decision about how many rows of data to extract at each point along the order, and we intend investigating the use of some optimal weighted extraction scheme. However, any such scheme needs accurate knowledge of the noise characteristics of the data and it would take considerable effort to implement a reliable automatic extraction algorithm.
Accordingly, the recommended approach for accurate extraction is first to use the `cdist' program to re-sample in the Y direction so as to straighten the orders. Experience shows that this program does an excellent job, with little discernible loss of resolution or variation of profile along the order. Once the orders are straightened, the number of rows to extract for object and for sky is merely a function of order number and not of wavelength.
FIGARO A general data reduction system