![](bg6.png)
STAR. Due to the efficient implementation the STAR
method reaches the runtime ranking two. In some
special sequences as Venus or Urban3 it gets good
covering performancewith a low feature set. But gen-
erally the performance of the STAR is below average.
SIFT. As shown in Figure 2 the SIFT detector run-
time depends on the number of features. It is the
slowest of the evaluation. But it shows good perfor-
mance within a low feature set. This method reaches
the best MEE results at Grove2 and Grove3 and the
best tracking performance at the difficult Urban2 and
Urban3. In our experiments the scale space analysis
of the DoG of the SIFT tends to detect less features
at object boundaries. Almost homogeneous regions
are also selected, e.g. the sky of the Grove3 sequence
where tracking is still applicable. Though the set of
objects are not covered uniformly, this results into low
coverage. Summarizing the SIFT detector is a slow
feature detector with a high tracking performance at
sequences with few objects.
SURF. In our experiments the improved runtime per-
formance regarding the SIFT could be confirmed, but
t
d
is still high compared to the other detectors. Re-
lated to SIFT the SURF shows some improvements at
large sets of features in terms of tracking performance
and accuracy. But FAST and GFT are still superior at
large feature sets.
MSER. The MSER is the only method based on seg-
mentation instead of finding features using first or
second order derivatives. In our experiments we could
observe that features are less likely detected at ob-
ject boundaries e.g. in the RubberWhale sequence
features are detected in the center of the fence holes
instead of their corners. Considering the generalized
aperture problem this is an important benefit. But it is
not reflected in the results of the whole dataset, where
the MSER detector achieve average results. Even at
the Venus sequence it performs the best MEE results.
This sequence is also a challenging sequence for
the GFT because it mainly consists of homogeneous
objects. Most of the cornerness regions are distributed
at the object boundaries. Thus MSER and also the
scale space detector SIFT and SURF achieve good re-
sults. That leads to the conclusion that a region based
or a scale space based approach would be attractive to
improve the GFT method.
6 CONCLUSIONS
In this paper we evaluate different feature detectors in
conjunction with the RLOF local optical flow method.
While the state-of-the-art KLT-Tracker operates with
the GFT detector, our goal is to provide a set of base-
line results for alternative feature detectors to allow
researchers to get a sense of which detector to use
for a good performance on their specific problem. To
compare the resulting sparse motion fields we pro-
pose a methodology based on accuracy, efficiencyand
coveringmeasurements. The benchmark is performed
on the Middleyburry dataset, which provides a set of
consecutively captured images and the corresponding
dense motion ground truth.
We observethat the efficient performingalgorithm
is the FAST approach. With a runtime of about 3 ms
for a VGA sequence and an overall good performance
for the tracking efficiency, median endpoint error and
coverage it outperforms the standard GFT detector.
For standard scenarios i.e. uniform texture distribu-
tion the GFT is a fast and reliable feature detector.
From the results given by the evaluations we conclude
that for specific sequences, e.g. sequences includ-
ing homogeneous objects, where the texture distribu-
tion concentrates on object boundaries, there is still
room for improvements. Limited advantages shows
the PGFT but also SIFT and MSER shows enhanced
performance under these conditions. Furthermore an
improved GFT method could benefit from the scale
space analysis of the SIFT or the region based ap-
proach of the MSER.
ACKNOWLEDGEMENTS
The research leading to these results has received
funding from the European Community’s FP7 under
grant agreement number 261743 (NoE VideoSense).
REFERENCES
Agrawal, M., Konolige, K., and Blas, M. (2008). Censure:
Center surround extremas for realtime feature detec-
tion and matching. In European Conference on Com-
puter Vision (ECCV 2008), pages 102–115.
Ali, S. and Shah, M. (2007). A lagrangian particle dynam-
ics approach for crowd flowsegmentation and stability
analysis. In Computer Vision and Pattern Recognition
(CVPR 2007), pages 1 –6.
Baker, S., Scharstein, D., Lewis, J., Roth, S., Black, M. J.,
and Szeliski, R. (2009). A database and evaluation
methodology for optical flow. Technical report MSR-
TR-2009-179, Microsoft Research.
Bay, H., Ess, A., Tuytelaars, T., and Gool, L. V. (2008).
Surf: Speeded up robust features. Computer Vision
and Image Understanding, 110(3):346–359.
Bouguet, J.-Y. (2000). Pyramidal implementation of the lu-
cas kanade feature tracker. Technical report, Intel Cor-
poration Microprocessor Research Lab.
ICPRAM 2012 - International Conference on Pattern Recognition Applications and Methods
308