GSI Forum: Tracking » Kalman low efficiency

Home » PANDA » PandaRoot » Tracking » Kalman low efficiency

Show: Today's Messages :: Polls :: Message Navigator

Kalman low efficiency [message #11946]

Thu, 02 June 2011 12:04

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *1-79-r.retail.telecomitalia.it

Dear all,
while analysing the data samples of single track events simulated for the STT/TPC comparison, we have noticed something strange: the reconstruction efficiency after the Kalman fit is much lower than the efficiency after the STT+MVD pattern recognition, especially for low momentum tracks.

In the following table, an example of the efficiencies of the STT pattern recognition, the STT+MVD PR and of the Kalman fit are reported:

Muons @ 0.3 GeV/c, theta = 40°
STT alone PR: 87%
STT+MVD PR: 97%
KALMAN : 65%

Muons @ 1 GeV/c, theta = 40°
STT alone PR: 87%
STT+MVD PR: 94%
KALMAN : 89%

The efficiency is calculated as (integral in peak)/(number of generated events), where the peak is defined as the range (mean-3sigma, mean+3sigma).
As you see, the high efficiency of the STT+MVD pattern recognition is reduced after the Kalman.

Two plots are attached to the message, showing the momentum distributions obtained with the STT alone (black), with the STT+MVD reconstruction (red) and after the Kalman (blue), again for muons at 0.3/1 GeV/c and theta = 40°.
As it is shown in the file "03GeV_40deg.pdf", the Kalman produces long tails in the momentum distribution, causing a lower efficiency in peak; these tails are not present in the other two distributions. In addition, the distribution is no more gaussian.
The tails are not present in the Kalman 1 GeV/c distribution; nevertheless, the efficiency is lower than the STT+MVD one also in this case.

Is there something going wrong in the Kalman?
Has someone already noticed this behaviour or has any idea/suggestion?
Is it possible for the detector experts to check if the recohit covariances are correct?

Thank you.
Ciao,
Susanna

Attachment: 03GeV_40deg.pdf
(Size: 16.27KB, Downloaded 299 times)
Attachment: 1GeV_40deg.pdf
(Size: 16.30KB, Downloaded 298 times)

Report message to a moderator

Re: Kalman low efficiency [message #11948 is a reply to message #11946]

Thu, 02 June 2011 12:19

Anonymous Poster

From: *adsl.alicedsl.de

Hi Susanna,

in the tracks that enter into the plot for the Kalman filter, do you check that the getStatusFlag() of the track rep is giving 0?

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #11949 is a reply to message #11948]

Thu, 02 June 2011 12:31

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *1-79-r.retail.telecomitalia.it

Hi Christian,
I check the flag of the Kalman output track with GetFlag().
If the flag is <0, then the track is not included in the plot.
Isn't it enough?

Ciao,
Susanna

Report message to a moderator

Re: Kalman low efficiency [message #11951 is a reply to message #11949]

Thu, 02 June 2011 12:33

Anonymous Poster

From: *adsl.alicedsl.de

Hi,

what is the call and check you make, exactly?

Christian

Report message to a moderator

Re: Kalman low efficiency [message #11953 is a reply to message #11951]

Thu, 02 June 2011 13:07

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *1-79-r.retail.telecomitalia.it

Hi,
for each event, I do the following:

    for (Int_t k = 0; k < kalfit->GetEntriesFast(); k++) {
       kalTrack = (PndTrack*) kalfit->At(k);
       if (!kalTrack) continue;
       if (kalTrack->GetFlag() < 0) continue;

where kalfit is the TClonesArray of PndTracks produced by the Kalman:

      TClonesArray *kalfit = new TClonesArray("PndTrack");
      reco->SetBranchAddress("SttMvdGenTrack", &kalfit);

Then, if the track is a primary track (I check the MCTrackId of the corresponding TrackCand), I fill the histogram of the momentum distribution:

     hgen->Fill(kalTrack->GetParamFirst().GetMomentum().Mag());

Susanna

Report message to a moderator

Re: Kalman low efficiency [message #11954 is a reply to message #11953]

Thu, 02 June 2011 13:10

Anonymous Poster

From: *adsl.alicedsl.de

Hi,

I see. Could you please check how this PnTrack::GetFlag() relates to the GFTrack::getStatusFlag() of GENFIT?

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #11955 is a reply to message #11954]

Thu, 02 June 2011 15:51

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *13-87-r.retail.telecomitalia.it

Hi,
I had a look into the code.
The PndTrack flag can be filled both in PndRecoKalmanFit and in PndGenfitAdapters, with the following values:
i) -10 if pz=1e-9
ii) -2 if the conversion from GenfitTrack to PndTrack has failed
iii) -1 if the the number of degrees of freedom (NDF) = 0
iv) 1 if NDF != 0

Concerning fStatusFlag, it's a data member of GFAbsTrackRep. I see that in GFKalman it is set to 1 if you catch an exception in processHits(); then there is a continue and you go to the next rep.

So if I understand correctly, the values of these two flags do not seem to be related. Or am I wrong?

Susanna

Report message to a moderator

Re: Kalman low efficiency [message #11956 is a reply to message #11955]

Thu, 02 June 2011 16:14

StefanoSpataro
Messages: 2736
Registered: June 2005
Location: Torino

first-grade participant

From: 222.190.126*

In the original PndGenfitAdapters::GenfitTrack2PndTrack (Christian was the author) there was no flag propagation from genfit track to PndTrack.
The flags Susanna is speaking about where put by me in order to check if something had failed in the conversion procedure and to fix eventual bugs.
Therefore, reading the code, GFTrack::getStatusFlag() is not accessible from PndTrack at the moment.

Report message to a moderator

Re: Kalman low efficiency [message #11958 is a reply to message #11956]

Thu, 02 June 2011 17:28

Anonymous Poster

From: *adsl.alicedsl.de

Hi Susanna,

so you you please add GENFIT flag to the PndTrack? The important thing is: status==0 means everything is fine, status!=0 fit failed. Then please apply a cut on this flag and redo the momentum spectrum. I am wondering if the tails get smaller...

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #11959 is a reply to message #11958]

Thu, 02 June 2011 18:19

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *13-87-r.retail.telecomitalia.it

Yes, I can try but not immediately because other tests are running on my machine and I cannot change the code meanwhile.

Anyway, I fear that applying a cut on this flag will influence only the distribution tails (in case) and will not affect at all the problem of the efficiency lowered by the Kalman. Do you have any idea about that?

Ciao,
Susanna

Report message to a moderator

Re: Kalman low efficiency [message #11960 is a reply to message #11959]

Thu, 02 June 2011 18:34

Anonymous Poster

From: *adsl.alicedsl.de

Hi,
at the end of the day we need to know:
-What fraction of tracks can be correctly fitted with GENFIT?
-Why are the tracks that are not fitted not fitted?
-Is it maybe due to false hits picked up in the pattern reco?

The last point you can check by looking at the efficiency after a MC truth pattern reco. If you implement the GEFIT flag, just make sure that the other cases the PndTrack flag can take (pz==1e-9, etc) do not superseed the GENFIT flag. When we isolate for which tracks GENFIT fails, we can try to find out why.

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #11967 is a reply to message #11960]

Mon, 06 June 2011 18:29

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *pv.infn.it

Hi Christian,
I made the test on the Genfit flag and I found that only in 2 events (out of 10000) the GF status flag is equal to 1 (failure).
So this should not be the reason of the tails...
Concerning the other failures of Genfit, by looking at the flag implemented by Stefano, I found:
- 46 events with flag = -1 (NDF=0)
- 310 events with flag = -2 (GFTrack to PndTrack conversion failed).
As I told you, these events are rejected and do not enter in the plot.

Do you have any suggestion?

Ciao,
Susanna

Report message to a moderator

Re: Kalman low efficiency [message #11998 is a reply to message #11967]

Mon, 13 June 2011 10:21

Anonymous Poster

From: *dynamic.mnet-online.de

Hi Susanna,

sorry for not replying earlier, I was out of office...

I see, the GENFIT flag does not explain things. There is one more thing that just came to my mind:
What values for the momentum and position initialization due you use for building the GeaneTrackRep objects that go into GENFIT?

I ask this because I remember a similar behavior in cases where the init pos and mom values were not very good in some test I did in the past. What was happening was that I was basically falling into the left-right ambiguity normal for drift chambers.

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #12001 is a reply to message #11998]

Tue, 14 June 2011 11:33

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *pv.infn.it

Hi Christian,
the GeaneTrackRep that goes into Genfit is built with the PndTrack that comes out from the STT+MVD pattern recognition, then backpropagated to the point (0,0,0).
The momentum distribution of the PndTracks after the pattern recognition is the red one shown in the plots attached to the first message: it seems reasonable and without tails.
Then, before going into Genfit, the tracks are backpropagated to (0,0,0) but I don't have any plot for these.

Ciao,
Susanna

Report message to a moderator

Re: Kalman low efficiency [message #12098 is a reply to message #12001]

Mon, 20 June 2011 14:13

Anonymous Poster

From: *adsl.alicedsl.de

Hi,

could you try as a test to initialize the GeaneTrackRep with MC truth values for the position and the momentum. This way we can exclude that there are problems with the seed values for the Kalman fit.

Cheers, Christian

Report message to a moderator

Re: Kalman low efficiency [message #12177 is a reply to message #12098]

Thu, 30 June 2011 10:17

Anonymous Poster

From: *adsl.alicedsl.de

Hi Susanna,

did you find out more about this problem? Did the seed values have anything to do with your problem?

Bye, Christian

Report message to a moderator

Re: Kalman low efficiency [message #12178 is a reply to message #12177]

Thu, 30 June 2011 10:51

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *pv.infn.it

Hi Christian,
unfortunately I have problems with the latest version of the pandaroot, so I couldn't try the test you suggested yet.
As soon as I can perform the test, I will let you know the results.
Ciao,
Susanna

Report message to a moderator

Re: Kalman low efficiency [message #12222 is a reply to message #12177]

Wed, 06 July 2011 12:51

Susanna Costanza
Messages: 33
Registered: January 2008
Location: Pavia

continuous participant

From: *pv.infn.it

Hi Christian,
I made the test you suggested, initializing the GeaneTrackRep with the MC position and momentum values.
The results are in the plot attached: the red histogram is the input one, with the MC values; the blue one is the Kalman output.
Also with the true values as input, the results are not good...
Any suggestion?

Ciao,
Susanna

Attachment: 300MeV_MCseeds.pdf
(Size: 14.14KB, Downloaded 256 times)

Report message to a moderator

Previous Topic:	MC true matching to reconstucted one
Next Topic:	TPC signal data available on GRID

Goto Forum:

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

] [

PDF

]

Current Time: Mon Dec 02 20:53:16 CET 2024

Total time taken to generate the page: 0.00804 seconds