In the following, some guidelines for a consistent performance evaluation
are collected.
A second ingredient is the geometrical range of the detector. It does not make sense to judge the performance of a reconstruction program on tracks which which are outside the acceptance of the detector, or just straddling its border. Objects of interest within the geometrical acceptance region define the reference set.
In many cases, the geometrical criterion can be conveniently defined
by the number of detector elements, eg. the number of layers in a tracking
system, which have been passed by the particle, which can be derived from
the Monte Carlo Impact Points (MIMPs) provided by the detector simulation.
Example: for the vertex detector, tracks within the acceptance region pass
at least three superlayers. Assuming that each superlayer has four logical
GEDE planes on two detector modules per quadrant, giving rise to two MIMPs
per superlayer, a requirement of at least 6 MIMPs for a reference particle
may be appropriate. It is permissible to discard MIMPs outside the sensitive
region from this count but it should be stated.
eXgeo = NXref / NXtotwhich is independent of the hit efficiencies of the individual detectors. Since the definition of the reference set has potential influence on the reconstruction efficiency, the corresponding geometrical acceptance should always be considered as well.
Parameter matching is sometimes the only way because it requires not direct Monte Carlo relation information between generated particle and the detector hits used in the reconstruction. In high particle densities however, it bears the danger to accept random coincidences between Monte Carlo particles and ghost reconstructions and can lead to the paradox impression that the reconstruction efficiency improves with increasing hit density.
- parameter matching
- hit matching
The recommended technique for HERA-B is therefore based on hit matching:
the reconstructed track is assigned to the Monte Carlo particle
which contributes the largest fraction of its hits if this fraction exceeds
a certain limit. The contribution fraction of a Monte Carlo particle m
is defined as
fm = Nmhits / Nhitswhere Nhits is the total number of detector hits on the reconstructed track and Nmhits the number of detector hits to which the MC particle m has contributed. Note that since several MC particles may contribute to the same strip cluster etc, the sum over all fm for a reconstructed track can in general exceed unity.
The track is assigned to the particle of largest contributed fraction if it exceeds a lower limit fmin. A typical choice for tracking applications is fmin=70%.
A particle is called reconstructed if at least one reconstructed object
is assigned to it. The reconstruction efficiency is then
e = Nrefreco / Nrefie. the fraction of reference tracks which are reconstructed.
A ghost rate can be defined as
egh = Nghost / NrefIt is informative to also provide the mean number of ghosts per event.
Nmclone = Nmreco - 1 , if Nmreco >0and one can define the clone rate as
0 , otherwise
eclone = Nclone / Nref
The quality of the parameter estimate is reflected in the parameter
residual
R(Xi) = XiREC-XiMCwhere Xi is a reconstructed particle parameter (eg. x, y, tx, ty, Q/p, ...). The corresponding Monte Carlo value can usually be inferred from the Arte tables MTRA,MIMP,MCAL etc. From the parameter residual distribution, one can then obtain the parameter estimate bias, <R(Xi)> , and the parameter resolution given by the width, which can be expressed in terms of the result of a Gaussian fit, the RMS or FWHM, whichever appears appropriate with regard to the shape of the distribution.
If the reconstruction tool provides also an estimate of the parameter covariance matrix (Cij), it is very informative to investigate the normalized parameter residual or pull, which is defined as
P(Xi) = (XiREC-XiMC) / (Cii)1/2Ideally, the pull should follow a Gaussian with mean value zero and standard deviation one.