Details
-
Type:
Improvement
-
Status: To Do
-
Resolution: Unresolved
-
Fix Version/s: None
-
Component/s: afw, meas_algorithms
-
Labels:None
-
Team:Data Release Production
Description
This is a request to create a mask bit indicating pixels that should be ignored for all purposes, including object detection. The idea for this is images with large contiguous regions that contain no useful data. Specifically, a region with this mask bit set would be ignored in the object detection stages.
A specific use case motivating this request is the use of the LSST DM software on co-adds from sets of individual dithers (potentially ISRed and combined outside DM stack). These co-adds have effectively circular regions of useful validity, and corners that should be ignored. In principle, the variance of these regions should also be set high, but the proper estimation of variance in regions of little data is not always robust (and quite frankly is often done wrong by the user – including this one). The FITS image array must be a rectangle, so the use of a mask bit (which can be visualized as a plane in the mask) to indicate that these regions should be ignored at all stages would be the clearest way of marking this.
Attachments
Issue Links
Activity
We'll probably need to turn this into an epic and schedule it in a future cycle.
Reviewing for CCB: I've long thought this was a good idea, but agreed with previous commenters that it'd be a lot of work, at least to do it directly. Thinking about it again, setting pixels within this mask plane to zero temporarily (we already use a temporary image in detection, since we apply a PSF-sized Gaussian filter to the original) would also ensure we never find anything above threshold, and should be quite easy to do in SourceDetectionTask.
Which mask planes to apply this logic to should of course be configurable, but the existing NO_DATA plane might be a good default; I'd need to remind myself if that gets used in any problematic cases first, like the cores of saturated-in-every-epoch stars on the coadd.
I recently looked at this on the HSC side, and left it alone because it would not be straightforward. I would like to see it happen, but it should involve a big cleanup of the detection code.
One workaround would be to set the image value of the masked pixels low, which is very easy in python.