Overview: Detecting Sources in Imaging Observations

A. Factors Affecting Performance And Runtimes

• celldetect

  The primary factor that influences the runtime of celldetect is the size of the dataset. Two secondary factors are the number of cell sizes to examine and the actual number of detected sources.

• wavdetect

  The three most important factors that affect the runtime of wavdetect are:

  1. the size of the dataset
  2. the number and size of wavelet scales
  3. the number of background cleaning iterations.

  The first two factors influence both wtransform and wrecon, the third only affects wtransform .

  For large wavelet sizes, the dataset needs to be padded. This causes the size of the data to be artificially augmented.

  wavdetect also uses a lot of computer memory. At present, 1024 x 1024 pixels should be considered the practical limit for the size of the input dataset on most machines.

• vtpdetect

  The runtime of vtpdetect does not depend on the spatial size of the dataset, but depends on many other factors and is rather unpredictable. The most important factors are the number of unique photon locations and the overall number of events.

  vtpdetect will run quickly if the number of photons is low and there is a high contrast between background and sources. In the opposite scenario (i.e. when one has a large number of photons and a large number of faint sources), the vtpdetect run can be very long, since fitting background becomes an arduous task. In these situations, our tests have shown that vtpdetect becomes very slow if the observation contains more than ~10⁵ photons.

---

B. Sample Runtimes

To give the reader a rough idea on the performance, we ran all three tools on various subsets of two simulated HRC-I observations. One simulation contained a set of sources on top of a flat background, the other contained only a flat background, but for a 10 times longer exposure, i.e. containing roughly 10 times more background events.

The following table shows the runtimes on a Sun Ultra 1 with a 167 MHz processor and 124 MB of RAM. The same runs on a Sun SPARCstation 5 with 70 MHz CPU and 64 MB of RAM were ~3 times slower. The runtimes of celldetect and wavdetect are not affected by the increase in the number of events; the major factor is instead the size of the dataset. In the case of vtpdetect, however, the number of events is the primary factor. The tool slows down considerably if the number of events is large.

Some things to keep in mind about the runs reported in the table:

• only one wavelet scale was analyzed for wavdetect
• only one background iteration was performed for both wavdetect and vtpdetect

On account of these two items, actual runtimes will normally be longer than indicated. Also:

• to a first approximation, the time required for execution of wrecon and vtpdetect increases linearly with the number of iterations.
• each additional scale in wavdetect increases the runtime by roughly the same amount of time.
• The ``recursive blocking scheme" was used in celldetect runs for 32k x 32k data.

Size (pixels)	Nevt	Approximate run time (min:sec)
		celldetect	wavdetect	vtpdetect
Short exposure, bkg+sources
512 x 512	400	0:25	3:45 (3:00+0:45)	0:03
1024 x 1024	1300	0:50	11:40 (9:30+2:10)	0:05
2048 x 2048	3800	2:00	--	0:16
32768 x 32768	1.1e105	8:50	--	9:40
Long exposure, bkg only
512 x 512	1800	0:25	3:35 (2:55+0:40)	0:15
1024 x 1024	7300	0:50	11:30 (9:30+2:00)	0:50
2048 x 2048	28000	2:00	--	1:40
32768 x 32768	1.1e106	8:50	--	--

---

C. Comparative Results For The Chandra Deep Field South

As described in section 2.4 of the Detect manual, we chose to use ObsID 2405 (part of Chandra Deep Field South) as a test field because it contributes 10% of the total exposure time of the Deep Field South. From running detect on the full exposure, we know the locations of all real sources which could be detected from the short exposure.

For this comparison, we have run the tools with their default parameter values. For celldetect, we used a S/N threshold of 3, an encircled energy of 0.9, and the PSF library of 2001 April. For wavdetect, we used scales of 1,2,4,8, & 16 and sigthresh = E-6. We found 71 detections for the block=2 image and 26 for the inner 1024 x 1024 area at block=1; of these, only 4 were unique. For vtpdetect, we used scale = 1 and a maximum probability of being a false source (limit) equal to E-5 and E-6. The results are shown in the following table.

Tool	# of Detections	# of Edge Detections	# of Spurious Detections	# of Valid Detections	Percent Spurious
celldetect	36	1	1	34	3%
wavdetect	75	19	0	56	0%
vtpdetect (E-5)	61	0	6	55	10%
vtpdetect (E-6)	48	0	2	46	4%

The distinction between "spurious" and "edge" is somewhat subjective. In the case of wavdetect , the edge detections were close to, but not actually on, the edge. The percent spurious assumes that these detections don't count; i.e. if we had used an exposure map, they probably would not have been detected. By adjusting various parameters, each tool could be made to perform better or worse; the values presented here are indicative of the norm.