Details
-
Type:
Bug
-
Status: Done
-
Resolution: Done
-
Fix Version/s: None
-
Component/s: ap_pipe
-
Labels:
-
Story Points:8
-
Epic Link:
-
Team:Alert Production
Description
As first identified in DM-13081, the template mask planes seem to propagate into unreasonably large masks in the difference image. This ticket is to investigate the source of this behavior and fix it, if appropriate. It may be useful to consult with Gabor Kovacs [X].
Attachments
Issue Links
Activity
Ok, I will do debugger runs anyway, so I'll try to identify the relevant codes modifying the masks.
Gabor Kovacs [X], Eric Bellm — now that we have a better understanding of ip_diffim in general, do we still regard this ticket as an issue?
(I'm assuming that means no).
Gabor Kovacs [X] reports that this behavior is due to the mask planes being OR'd during kernel convolutions, even for pixels where the kernel weights are 0.
The relevant declaration in include/lsst/afw/image/Pixel.h:73
template <typename _ImagePixelT, typename _MaskPixelT, typename _VariancePixelT = double>
class SinglePixel : public detail::MaskedImagePixel_tag {
public:
template <typename, typename, typename>
friend class Pixel;
template <typename T>
friend class PixelTypeTraits;
typedef _ImagePixelT ImagePixelT;
typedef _MaskPixelT MaskPixelT;
typedef _VariancePixelT VariancePixelT;
...
convolution multiplies and adds SinglePixel instances include/lsst/afw/math/ConvolveImage.h:274
inline typename OutImageT::SinglePixel convolveAtAPoint
...
typename OutImageT::SinglePixel outValue = 0;
...
outValue += *inImageLocator * kVal;
include/lsst/afw/image/Pixel.h:604
template <typename ExprT1, typename ExprT2>
BinaryExpr<ExprT1, ExprT2, std::multiplies<typename exprTraits<ExprT1>::ImagePixelT>,
bitwise_or<typename exprTraits<ExprT1>::MaskPixelT>,
variance_multiplies<typename exprTraits<ExprT1>::VariancePixelT> >
operator*(ExprT1 e1, ExprT2 e2) {
return BinaryExpr<ExprT1, ExprT2, std::multiplies<typename exprTraits<ExprT1>::ImagePixelT>,
bitwise_or<typename exprTraits<ExprT1>::MaskPixelT>,
variance_multiplies<typename exprTraits<ExprT1>::VariancePixelT> >(e1, e2);
}
For all (+,-,*,/) pixel operations, the mask operation is bitwise_or. Actually we do not have any operations with bitwise_and.
Proposal:
I'd propose a variant of the convolution operation that performs bitwise_and at the image and kernel pixel multiplication step. So that the kernel as a MaskedImage could carry a region "AND-mask" which would be bitwise_or-ed around a result pixel by the addition step. In this way, regions depending on the kernel value could be defined that determine the mask value of the output pixel.
Great! Per discussion of 2019-10-15, we should:
- Write another ticket to capture the work of RFCing the proposal above;
- Claim this ticket as “done”.
There is a if (kval != 0) condition around the pixel multiplications that I've overlooked earlier. Nevertheless, for AL kernel solutions the kernel pixels are never exactly zero. (They may for other kernel classes, where an analytic function was not evaluated, but I don't fully see the implementation details. )
Gabor Kovacs reports that this behavior is due to the mask planes being OR'd during kernel convolutions, even for pixels where the kernel weights are 0.
We (Meredith, Eric, John) discussed this on 2019-01-29 and thought it probably makes sense to assign this to Gabor, although we don't think it should be his immediate priority.