In computer science, "shepherding" is often essentially a means of adapting the journal Major Revision decision for conference publications. Because conferences normally have only one round of review, it is generally the case that the only two decisions available are "Minor Revision" (i.e., the authors are trusted to fix their paper without need for further review) or "Reject."
Shepherds allow a conference to accept papers that need more serious revision without endangering the quality standards of the conference. Rather than simply accept the paper, the conference gives it a conditional acceptance and appoints somebody from the program committee to work with the authors to ensure that the paper ends up acceptable (or to ultimately reject it if the authors cannot reach acceptable quality).
As such, thanking shepherds in acknowledgements often makes a lot of sense, since this is a person who has given volunteer time to work directly with the authors to ensure their paper is significantly improved from its original form.