Frequently Asked Questions about jitter

FAQ Revised: Friday 23 September 2005 12:12:00

Table of Contents

1. General
2. Input and configuration files
3. Sky estimation/subtraction
4. Shift and Add
5. Software-related questions
6. Output FITS file format

1. General

1.1. What is a FITS file?
See the FITS support office home page at NASA for more informations about this file format.

1.2. How do I reduce infrared data?
Contact an infrared expert to get training. Reducing infrared data is no trivial task. The ISAAC pipeline tools are trying to make your life easier by automating some of the tedious tasks involved, but they do not replace the know-how of a trained astronomer. Read books, ask around you, read the ISAAC manual, etc. A good idea is usually to get training on infrared observing before actually submitting your proposals, or at least make sure you can get in touch with somebody who can help you process your data.

1.3. Can you help me reduce my data?
Short answer: NO

Long answer: the eclipse team is dedicated to provide pipeline tools to the community to help them reduce their data. We are by no means an absolute reference about infrared data reduction.

If you happen to find a bug or have comments about the methods offered by the pipeline, you are welcome to contact us: eclipse-team AT eso DOT org.

If you have problems with your data, you are invited to contact the User Support Group in Garching to clarify points related to data acquisition.

If you have problems understanding what should be done on your data, you need expert assistance. The eclipse team can unfortunately not provide that level of service. The User Support Group might give you hints. Reading all ISAAC documentation might be a good idea too. See the ISAAC Data Reduction Guide (from the ISAAC or eclipse home page).

1.4. What is jitter doing exactly?
Read the manual accessible from the eclipse home page.

1.5. What is a pipeline recipe exactly?
This question is answered in the general eclipse FAQ.
See also question below: "I heard some horrible things about jitter, what is true?".

1.6. May I use it for optical data?
You can of course use this software on optical data. Actually, number of people have been using it as a shift-and-add tool for images taken in the visible bands. The most sensible thing to do in this case would be to disable the sky estimation in the ini file, but otherwise algorithms are unchanged.

1.7. I heard some horrible things about jitter, what is true?
jitter is a very complex tool that is aimed at performing the data reduction steps with as little user interaction as possible. In an ideal future, it should be able to modify itself its internal parameters depending on the input data, to get the best possible results. But that is still a dream! Today you need to first read the documentation carefully to be sure you understand correctly what the input parameters are. Then you need to take the time to study your own case and define a correct set of parameters for your problem. Then you can run jitter and see if the results you get are consistent. The best data reduction ever is almost never achieved on the first attempt. jitter is not (yet!) a magical tool that does everything for you.

jitter is trying to be a good tool, not a good worker. There is no replacement so far for the attention an astronomer can have on a data set, and the scientific appreciation of the results. Consider this software as a transparent box that automates many tasks for you but has to be tailored to your data sets to produce the most efficient results.


2. Input and configuration files

2.1. Are there examples of jitter.ini files?
Use the -g or --generate option. It will produce a default jitter.ini file in the current directory.

2.2. What are the default parameters valid for?
Section by section:

  • SkyEngine
  • Default parameters are set for a slowly varying sky, i.e. the sky can be considered constant in time over a range of 2*RejectHalfWidth+1 frames. If you are observing in J or H with a total exposure time of 1 second per frame, it is reasonable to set the half-width of the running filter to a high value. If you are observing in K or Ks and have a long exposure time per frame, do not use values above 2 or 3 as half-width, the assumption of a constant sky in time is not valid anymore.

  • Shift and add
  • Default behaviour is to look for cross-correlating objects automatically, locate a small number of objects around the center of the first frame, and refine the inputs given by the FITS headers. Frame shifts are by default performed by resampling using a hyperbolic tangent kernel for interpolation. The only average method is filtered, rejecting a user-specified number of min and max pixels on each time line before averaging.

  • Post-processing
  • Default is to apply subtraction of the median from each row, to take care of saturation effects. No image viewer is launched at the end by default.



2.3. How much memory do you recommend to have?
It depends on the amount of data you have to process. A typical SOFI or ISAAC batch of jitter files is usually made of about 60 frames. That translates to roughly 256 megs in input, thus about 512 megs total memory consumption during the process. If you can afford a gigabyte of RAM for such data sets, do it. If you have less memory, jitter will create its own swap pages as necessary to make the best use of what you have.

For your information: jitter has been reported to reduce such SOFI batches on a Solaris station with only 32 megs of RAM (and loads of disk space), without any problem. The process is considerably slower, but it runs fine. I would say that 32 megs in RAM is the minimal amount to get a decent working jitter process. If you have a slow machine or little RAM, launch your processes at night with an at or cron command and get the results in the morning.

Do not push it too far, though. The memory handling mechanism in eclipse is able to deal with far more memory than you truly have (RAM chips + swap space on disk) because of this internal swap page mechanism. But it cannot allocate infinite resources. You might still run out of memory in the following cases:

  • No more disk space available on E_TMPDIR.
  • Too many opened files at the same time.
  • Your memory consumption has reached the limit set by your shell configuration (limit datasize).
  • You have exhausted all possible pointers on a 32-bit machine (no kidding this has already happened).
  • The output file you are trying to create is bigger than the maximum size for this filesystem.
  • Your kernel has run out of resources (no more pointers, size of static structures reach a maximum, etc.).
  • ...

There are many things that might fail if you are trying to launch processing of huge data sets, and there is nothing that can be done to prevent that. If you really need to run jitter on very large numbers of frames, you are probably doing something wrong anyway. See below about combining several data sets together.

2.4. How do I create the list of input files?
The simplest way is to use 'ls'.

% ls -1 *.fits > framelist.ascii

Replace *.fits by the wildcards you need to describe the files you want to process, and framelist.ascii by any name you want, provided it is declared correctly in jitter.ini.

Notice that ls is used with the option -1 (minus one), which requests a list of file names on a single column output.

Beware!

If your ls command is aliased to some exotic combination of options (most people alias ls to ls -F), the results of the previous commands may append some funny characters to the file names. Simple example: if your files have executable permission, and you aliased ls -F, all file names will have stars (*) appended, which will prevent jitter from finding the files. Notice that you can de-alias a command by preceding it with a backslash, i.e.

% \ls -1 *.fits > myframelist

If you want to do jitter+offset reductions, you will need to declare in second column the file type for each input frame (see next question). Hint: use dfits and fitsort -d to create this column output. For ISAAC, the simplest is to do:

    % dfits *.fits | fitsort -d DPR.TYPE > framelist.ascii


2.5. Is it Ok to put several data sets in the same frame list?
Short answer: NO

Long answer:

Here is an example: you have observed the same field during several nights, getting in the end N data sets, each of them containing typically 30 to 60 frames. Since you want to get an image of the field that includes all of your observations, you are wondering if you could give all these files to jitter to magically re-combine them to a single image.

Well... this is not possible. Several reasons:

  1. The sky subtraction will not work. You will have in the input list some consecutive frames taken widely apart in time. The sky filtering algorithm assumes a smooth (if not continuous) sky variation among frames, this breaks this assumption.
  2. The offsets declared in the header will not match. The rule is to declare every first frame in every batch being at offset (0,0). Now if you observe the first frame in each batch, you will see that they are not all exactly aligned, particularly if two data sets have not been taken the same night. This will bias the offset inputs and might lead to whole data sets not cross-correlating.
  3. Whatever machine you are running jitter on will most likely run out of memory during the process. Just realize that you have given the program several hundred megabytes to process, with a total requested load of about 2 times the size of your input data. Now check out how much memory you have on your machine and be more realistic.


2.6. So how to I combine several data sets of the same field?
The simplest solution is to reduce each batch separately, since each batch is consistent with itself and jitter is made for that. You will get one reduced frame per batch. Now you can combine these together using again jitter, but this time you need to disable the sky estimation and provide your own frame offsets, or if you have an idea of the maximal offset range, activate the automatic offset finder with large enough search sizes.

Agreed, this means that your frames will get interpolated a second time, but this should not be an issue. First, each reduced frame has already a quite high signal to noise ratio, which ensures that re-sampling will not act randomly. Second, there are by definition no very high frequencies in astronomical images because the acquisition system has been designed so. Re-sampling is not likely to cause artefacts with well-behaving signals like these.

If you are really worried about re-sampling the frames a second time, you can always work on the offsets between all frames to make them consistent relative to the first one in the batch. You will have to find a way of subtracting the sky in each batch to produce correctly sky-subtracted frames. And then you can try putting all your frames into the same framelist... and hope that your machine does not crash due to memory/CPU load.

2.7. How do I declare OBJECT and SKY frames?
The input list format is the following: provide the input file names in first column, and optionally the file type in second column. Both columns are separated by blanks or tabs. The file type in second column identifies whether the frame is an OBJECT or a SKY. The definition is: if the word sky (case-insensitive) can be found in the character string identifying the file type, then the frame is a sky. Otherwise it is an object.

Examples

The following declares only objects:
(no second column -> no sky string -> no SKY plane)

--- begin frame1.fits frame2.fits frame3.fits frame4.fits --- end

The following declares frame2.fits and frame4.fits as skies: (their second column contains a sky string)

--- begin frame1.fits frame2.fits sky frame3.fits frame4.fits sky --- end

Same for the following:

--- begin frame1.fits OBJECT_FRAME frame2.fits SKY_FRAME frame3.fits OBJECT_FRAME frame4.fits SKY_FRAME --- end

Same for the following (notice that the second column is case-insensitive to identify 'sky')

--- begin frame1.fits NGC9999 frame2.fits NGC_SKY frame3.fits NGC9999 frame4.fits skyFrame --- end

With SOFI/ISAAC frames, you can get away with a simple:

    % dfits *.fits | fitsort -d DPR.TYPE > framelist.ascii



3. Sky estimation/subtraction

3.1. Why all these dark crowns in output around extended objects or crowded regions?
During sky background estimation, the aim is to remove all objects from the sky background, to get something which is as close as possible to the real sky value. If this rejection does not work too well, there will be star signal left in the background. This background is then subtracted from all input frames, which means a signal which is a little bit negative in case some object signal was left in the background.

This will happen for example if you have crowded regions. The jitter movements are not sufficient (most will end up on object signal). For some pixels, most observed signal will be object signal, not sky. In that case it becomes impossible to filter reliably the sky out from this signal.

Warning!

It is a terrible idea to observe a very crowded field or an extended object with a simple jitter pattern. Use instead jitter+offset, it has been designed for that purpose. Of course, you are then spending observation time to observe a blank sky, but it is the only reliable way to get rid of this strong signal in the infrared regime.

There have been many cases already of observers who did not want to loose any time observing such useless signals as the sky, and were totally unable to reduce their data. If you have not calibrated the sky background, you are left out with atmosphere models to estimate it, without any guarantee of result. You are advised not to use the 'jitter' command in such cases.

3.2. How good is the sky-subtraction?
The sky-subtraction algorithm has been carefully studied by Angela Iovino. Her results are presented in a report available from the eclipse Web site.

Have a look at jitter and photometry report

3.3. How do I submit new sky-subtraction methods?
On average, we receive one new suggestion about sky filtering methods per week (no kidding). On average, there is also one valid suggestion every thousand. If you really think you have the ultimate sky-filtering algorithm, it is recommended to go to your favourite programming environment and try out your method. If you can prove on a reasonable set of data that your method is better than the current one, pack everything into a decently written report and send it to us.

An algorithm does not have to be believed but demonstrated. Just as you would not publish a scientific paper where you only quote your opinions, you want to prove that your method is working in a given domain with some restrictions and specified parameters. If you just want to make suggestions, make sure they are properly documented and the result of a scientific study, not just your own belief.

The next step is then to try to convince us of your method. There are too many suggestions about this topic for us to examine unsufficiently argumented ones, so be convincing!

3.4. Where is the sky background estimation?
The sky background estimation is given in the output QC PAF file. Have a look at your jittered_results_qc.paf file to get these measurements. Notice that there is no option any more to convert these values to other units, you will have to do it yourself. The given values are the measured average sky background values during sky estimation/subtraction.

3.5. How is the sky estimated in jitter+offset mode?
To summarize: a batch of jitter+offset frames usually means a small number of object and sky frames (10-20 at most). jitter will create a single sky plane by taking the median average of all sky frames in input, then subtract it from all object frames.

This method is Ok to provide a quick-look on the data set (which is what the pipeline is meant to do), but insufficient in many cases. In K band, the sky sometimes varies too fast for this method to provide a reliable sky plane. You might want to produce several sky planes in this case, combining them with whatever 3d filter seems appropriate to get a good estimate. It is definitely recommended to measure sky background variations on the sky frames before deciding on a way to filter them.

Once you have correctly subtracted out the sky, you can give the frames to jitter as you would with raw frames, disabling of course the sky estimation feature.


4. Shift and Add

4.1. How do I create an offset list?
Everything you need to know about offset lists can be found by launching jitter with the --offset option.

Here is one way to create simple offset lists. The easiest is to create it from header information. Using dfits, fitsort and awk you could do something like:

% dfits *.fits | fitsort -d seq.cumoffsetx seq.cumoffsety

That prints out the values of SEQ.CUMOFFSETX and SEQ.CUMOFFSETY for all FITS files in the current directory. Now if you want to create an offset list from this, you also need to add the plane number, which is usually just sequential from 1 to NPlanes. Using awk, you can easily add that by doing:

% [...] | awk '{print NR-1,$2,$3}' > offsets.in

[...] stands for the previous command, combination of dfits and fitsort.

4.2. jitter fails to find any object for cross-correlation. What should I do?
Try the following tricks:
  • Reducing the value of AutoThreshold will increase detection. Careful: decreasing this value too much will end up detecting noise and consequently increase the risk of false matches.
  • It might be that no bright object is situated within the inside box in the first frame. To overcome that, you can try to increase the size of the possible box by decreasing the cross-correlation search area or measure area. If your input offsets are coming from the FITS header, reduce in the [ShiftAndAdd] section the following values:
    • OffsetSearchSizeX
    • OffsetSearchSizeY
    • OffsetMeasureSizeX
    • OffsetMeasureSizeY

Last resort: if 'jitter' is still unable to find anything worth cross-correlating in your images, of if the objects it picks are not suitable for alignment (e.g. moving objects), you can provide your own list of objects in an ASCII file. See the corresponding sections in the jitter.ini file.

4.3. What if I am missing the first frame?
This should not change anything. jitter will identify the offsets between frames anyway, and the final reconstructed frame will be a union of all input frames (i.e. a bigger image than the input images).

4.4. What is the meaning of the returned value in parentheses, after the offset values?
The typical output from the offset matching routine is something like:

from plane 1 to plane 18:       [-4.24  -54.42] (25.42)
from plane 1 to plane 19:       [-57.44 -28.21] (33.38)
from plane 1 to plane 20:       [ 1.96  45.27]  (1396.27)
from plane 1 to plane 21:       [-18.35 -27.00] (35.13)
from plane 1 to plane 22:       [-37.79 -44.23] (1396.86)
from plane 1 to plane 23:       [37.00  41.88]  (25.61)

The 'jitter' command determines offsets between frames by using a cross-correlation criterion. This measurement is searching for an offset between the reference and the candidate frame, that minimizes the sum of squared differences between the frames. The value printed out in parentheses along with the offset measurements is the minimal difference value found in cross-correlation. The smaller the value, the better the match between frames. The theoretical limit for this value is zero (perfect match) in the case of a frame that is shifted by an integer translation vector. There is no absolute reference to measure the "goodness of match" with this value, since it will depend on:

  • The fractional displacement amount: it will increase the difference as displacements are closer to a half-pixel, and decrease as they get closer to integer pixels.
  • The noise present in the image.
  • The amount of signal in the image.

To have a visual indication of what this value is, you can have a look at the jitter algorithmic manual: http://www.eso.org/projects/dfs/papers/jitter99/node11.html

The figure shows what a cross-correlation matrix looks like. The value returned in parentheses by the 'jitter' command is the lowest value of this matrix (the minimum of difference).

As a rule of thumb, the values you get should all be more or less consistent. The lower the better, but the most important is to get variations in this distance measurement as small as possible.

4.5. What kind of errors should be expected due to shift-and-add?
There are several sources of errors in the whole process. Let us review them one by one.

Cross-correlation is failing on a number of signals. If your image exhibits a strong periodical pattern, any replica in the pattern might be a good candidate for cross-correlation, introducing large matching errors in the output. A solution is to increase the size of the measurement area until the reference zone does not exhibit any periodical pattern.

Another source of errors could be that the object used for cross-correlation reference is elongated (e.g. a galaxy). In the direction of elongation, any position is as likely to be picked as another. This would introduce systematic errors like an elongation of all stars in the final image. If you happen to be in such a case, try picking yourself the main object used for correlation, e.g. a bright star that can be seen on all frames. You could also increase the number of objects used in cross-correlation to compensate errors.

Resampling is not completely harmless. A new signal is derived from the input signal, using some assumptions in the process like signal smoothness. If you happen to have very high frequencies (typically bad pixels) this is likely to disturb the neighbouring pixels and introduce some errors in final pixel values. If your signal behaves correctly and smooth enough, you should not have introduced any noise at all through resampling, though.

The final stacking of images is also applying a 3d filter to remove the highest and lowest pixel values before summing up. This will systematically remove faint signals that appear on few frames, which might not be what you want to do. If you are interesting in keeping all input signal in the output, you should try linear averaging for the stacking and filter out bad pixels by yourself. Be aware that removing pixels systematically might be a cause of bias if your signal behaves irregularly.

4.6. Which method is used to shift the images?
The images are shifted using a high-order 2d separable interpolation algorithm. Integer shifts are to be avoided because of the aliasing they create.

There is an article about image resampling, containing a fair description of what the interpolation scheme is. You should have a look at the ESO Messenger (issue 100 - July 2000). The article is called: Astronomical image resampling. You can download the July 2000 issue of the Messenger from the ESO Web site, at:

http://www.eso.org/gen-fac/pubs/messenger/

4.7. What is the 3d averaging algorithm?
The final stacking is done by sorting out all pixels belonging to the same time line (z axis on the final cube), removing a user-specified number of min and max pixels, and averaging the rest. jitter used to offer various algorithms to do that, but this one is definitely the best. If you want to use a simple linear stack algorithm, you can set the reject numbers to 0,0. If you want to use a median filter, you can set the reject numbers so that only the middle value is kept.

4.8. Why is the final frame bigger than any input frame?
This change occurred in eclipse version 4. The default behaviour for the command is to try to reconstruct a union of all input frames. If you have small offsets between frames, this means that the final stacked image will be 10 to 15% bigger than the input frames. If you have large offsets between frames, the final image will grow similarly.

4.9. How are edge effects taken into account?
Pixels located on the edges of the final stacked image are by definition inconsistent with pixels located in the center, because they have seen less signal. To cope with this inconsistency, there are two methods:

  • Divide the final frame by the total number of input frames.
  • Divide each pixel in the final frame by the number of contributing pixels.

The first method maintains a consistent signal-to-noise ratio over the frame, the second one maintains a consistent photometry for objects located on the edges. Version 3 and earlier of eclipse offered only the first method. Version 4 currently only offers the second method, an option might be added later on if the need arises.


5. Software-related questions

5.1. jitter uses more RAM than I have!
Since version 3.0, eclipse uses a new (hopefully portable) memory allocation scheme. Without going into too many details, the library opens its own swap files in a temporary directory and tells the Operating System to use it as virtual memory.

Some process monitoring tools such as top will report these temporary files (of litterally unlimited size) as memory use, thus providing a false information. Do not be alarmed by the quantities announced on top on systems such as Solaris and Linux, they are just bugged. jitter has been reported to show a memory use of more than 2 gigabytes (on a 32-bit system!) with no more than half a gig of memory space (RAM chips + swap space)... Do not always believe blindly your memory monitoring tools.

5.2. Is there a way to save intermediate data in case of machine crashes?
No.


6. Output FITS file format

6.1. What is the meaning of the EXPTIME keyword?
According to the dictionary definition for ESO keywords:

EXPTIME provides the total integration time in seconds; it may have decimals. When the exposure is made of several periods, EXPTIME time is the sum of the exposure periods, and not simply the difference between end and start of exposure. Subintegrations, i.e. multiple exposures before a readout of the detector are described by the DIT and NDIT.

The reference is The Data Interface Control Board document.

This being said, the EXPTIME written in the output header by 'jitter' is the EXPTIME found in the first input header (NOT the sum of all EXPTIME values from all input files). This has been done upon request from the instrument scientist.

6.2. What are these HIERARCH ESO PRO keywords?
They are added to identify the frame as a pipeline product, in the ESO DataFlow environment. The only keywords you might be interested in are:

PRO CATG         product category, identifies the frame as COADDED_IMG
PRO DATE         date the jitter process was executed
PRO DATANCOM     how many frames were combined to produce this one