Image and Vision Computing

Dear Mr Zimmermann,

We have received reports from the Reviewers who were asked to review
your manuscript.  Please find the relevant parts of each review
attached to this email.  If after considering the review comments you
find you are in agreement with them you might wish to revise
your manuscript accordingly.

We would ask you to provide a written statement of the way you have
taken account of the referee comments in preparing the revision.  You
may include the text of this statement at the beginning or end of the
manuscript text when you submit the revised manuscript.
(It can be separated at a later stage).   

Please submit your revision by the Mar 31, 2009.

For your guidance, reviewers' comments are appended below.

To submit a revision, go to http://ees.elsevier.com/imavis/ and log in
as an Author.  You will see a menu item call Submission Needing
Revision.  You will find your submission record there. 

Yours sincerely

Keith Baker
Editor-in-Chief
Image and Vision Computing

Reviewers' comments:


Reviewer #1: Overview
The paper describes a system of tracking using a set of sequences of
linear predictors SLLiPs.
Rather than use a single linear predictor to estimate motion, a
sequence is learnt,
each of which refines the estimate given by the one before.  The basic
approach is
described in a PAMI paper [11].The contribution of this paper is to
describe
a method of learning the SSLiPs efficiently, in a way that is useful
for online
applications.  The performance (both in terms of timings and accuracy)
of the resulting algorithms and trackers is evaluated on a large test
set.

Methodology
The core of the paper is the training algorithm, which seeks to select
a sequence of
linear predictors which achieve a given predicted accuracy with as low
a computational
cost as possible.  This is a difficult combinatoric optimisation, so a
branch-and-bound
search is used.  The approach has the advantage of quickly producing a
tracker which
achieves the desired accuracy, but may not be the most efficient
possible.  This can
then be used to track an object in a real-time application, whilst the
search for
a better (more efficient) tracker proceeds in the background. 

The basic concepts are introduced in section 2.  The details of the
new learning algorithm are presented in 3.2.  However, they are
described so concisely that it is difficult to work out the exact
details of the algorithm.  Although the reader gets a good overview
of the approach,
it would be helpful if key details were expanded.

For instance, on p7 the idea of "restriction" is introduced - that of
generating a training set for a tracker of lower complexity by simply
missing out rows of the training set for a more complex tracker.
Clearly the number of ways in which this can be done is very
large - some description as to how one deals with this combinatorial
explosion would be helpful.

Similarly Fig 5 could do with more explanation.  Presumably when each
line is extended with two other lines, those correspond to one from
each of the two complexities considered.  However, it if choosing the
complexity also involves selecting (in advance) the subset
of pixels to be used in the tracker, or whether that is something that
is learned during the training process at the same time - presumably
not, as that would cause a combinatorial explosion.

The evaluation of the method is very good - a large dataset was used
and a range of methods compared.

Overall, this is a good paper, but it should be expanded to give a
more detailed description of the core algorithm.


Reviewer #2: Overview

The paper deals with the design of fast trackers, and proposes an
improvement to a tracker that consists of a Sequence of LLiPs
(SLLiP). Since the time required for motion prediction by a SLLiP
directly corresponds to the number of used pixels, it is desirable to
use
only a small subset of pixels from a template. 
The SLLiP learning is formulated as an optimization problem where time
of tracking (computational complexity) is minimized given a predefined
precision of motion predictors.
The learning might still be prohibitively time consuming for large
problems, while the intention is that the the method should provide a
real-time procedure for tracking.
The major contribution of the paper is describe an improvement in the
form of a learning approach which, after a very short initialization
period, provides a solution with predefined precision.

The performance tested upon a large data set compares very nicely to
other published methods.


The details of the new learning algorithm which is the major issue,
presented in section 3.2, should be extended further to facilitate
reading and understanding. 

Further details on the cases and reasons for failure of tracking would
be interesting and useful.

Details:

Spelling should be checked again to remove some spelling errors

Part of sentence on top of page 3 should read:

.. finds a globally optimal solution ...  OR
.. finds the globally optimal solution ...  OR
.. finds globally optimal solutions ...