Changes between Version 4 and Version 5 of ppSub_vs_Hotpants

Timestamp:

Jan 19, 2010, 5:34:12 PM (17 years ago)

Author:

Paul Price

Comment:

Adding update: large kernels make better subtraction.

ppSub_vs_Hotpants

@@ -100 +100 @@
 
 ppSub (without adding photometry) has a running time about twice that of hotpants, but it is performing four times as many convolutions.  This better run time per convolution is probably due to ppSub doing convolutions via Fast Fourier Transforms, rather than directly.
+
+= Update =
+
+The following parameter variations result in much cleaner subtractions:
+
+{{{
+KERNEL.SIZE: 40 (Half-size of kernel)
+STAMP.FOOTPRINT: 40 (Half-size of stamps)
+ISIS.WIDTHS: 1.7 2.9 4.8 8.2 13.8 17.9 (FWHM of Gaussians)
+ISIS.ORDERS: 4   4   2   2   2    0  (Polynomial order for Gaussians)
+}}}
+
+[[Image(source:/trunk/doc/ppsub/pics/comparison/isis_radial_2.jpg)]]
+
+The widths of the bases are given by 1.69^i^; the idea is to try to keep the widths from being integer multiples of each other.  We also go up to a large width to support the largest kernels.  Unfortunately, this requires very large kernels, which makes finding stamps very difficult given the maskiness of GPC1.  Work is under way on scaling these parameters according to the inputs.
+
+Some sources are still detected in these images.  29-30 is the worst, with 94 sources.  Note that 29-30 is approaching worst case: very elongated (AR = 1.4) and cross-directed (theta = -15,+105) PSFs.  Other cases with more than 10 sources detected are:
+
+{{{
+28-30: 84
+28-31: 81
+28-32: 28
+28-36: 48
+29-30: 94
+29-31: 91
+29-32: 27
+29-36: 49
+30-32: 35
+30-35: 22
+31-32: 33
+31-35: 16
+35-36: 52
+}}}