ctools

• Smoothing of the input map
• Bin-independent computation of mean and standard deviation taking into consideration only those pixels within a certain distance. This is done using a convolution with a disk kernel.

• Old definition: Pixel with the highest value in the map
• New definition: Pixel with the largest number of standard deviations above its respective mean.

The two definitions are the same if the radius used for computing the mean and stddev. is set to cover the entire skymap.

• smoothkernel: Kernel to use for pre-smoothing the skymap (possible values are NONE, GAUSSIAN and DISK)
• smoothparam: Parameter for the smoothing kernel (sigma for GAUSSIAN or radius for DISK, ignored if smoothkernel=NONE)
• momentradius: All pixels within this distance of a given pixel will be used for computing its respective mean and standard deviation.

Comparing the computation time between the previous method and the updated method are as follows (note that the lack of smoothing does not dramatically impact the computation time):
Parameters:

inmap= counts map of the GPS survey (7200 x 280 pixels)
maxsrcs=100
threshold=3

Devel version:

real 19m14.410s
user 19m0.748s

2498 changes (smoothkernel=NONE, momentradius=1 deg):

real 26m4.866s
user 25m37.019s

2498 changes (smoothkernel=NONE, momentradius=5 deg):

real 26m45.997s
user 26m13.691s

2498 changes (smoothkernel=NONE, momentradius=10 deg):

real 26m18.184s
user 25m42.642s

Here’s an image comparing the results for the two methods showing the detected source positions for the devel branch (green crosses) and the changes branch with momentradius=5 deg (magenta diamonds) and 10 deg (yellow circles). There appear to still be some missed sources, but these can probably be recovered by increasing the value of maxsrcs.

In addition to the above changes, I’ve also taken the liberty to adapt the GSkyDir::cos_dist to speedup the computation in 'cssrcdetect._remove_maximum()'. This saves about 2 minutes (~7.7%) in the overall computation time. I’ve also cached the sky positions for the 'find_maximum()' and '_remove_maximum()' methods. This yields an additional 9 minutes (~37%) reduction in computation time, taking just over 15 minutes.

I’ve attached the resulting region and model files, in case that is of interest.

Pull branch:
Josh Cardenzana/ctools: 2498-cssrcdetect_local_selection

I’m about to merge in your changes. In which GammaLib branch are the changes in GSkyDir::cos_dist?

I’m also wondering whether we should switch the new parameters to hidden, at least the ones for smoothing. They all affect more the inner workings of the source detection and are there to tune the result. What about the following?

momentradius = 1.0
smoothkernel = "GAUSSIAN" 
smoothparam = 0.05

Here is finally what I implemented

avgrad,    r, h, 1.0,,, "Averaging radius (deg) for significance computation" 
corr_rad,  r, a, 0.1,,, "Correlation kernel radius (deg)" 
corr_kern, s, h, DISK,NONE|DISK|GAUSSIAN,, "Correlation kernel to apply to input sky map" 

Code is integrated in devel branch.