Paper: 28
Title: Simultaneous learning of motion and appearance


-------------------- review 1 --------------------

PAPER: 28
TITLE: Simultaneous learning of motion and appearance

OVERALL RATING: 1 (weak accept)
REVIEWER'S CONFIDENCE: 2 (medium)
----------------------- REVIEW --------------------

This paper presents a method for simultaneous learning of motion and appearance
parameters for object tracking. The parameters for the object appearance are
estimated using a Learned Linear Predictor (LLiP). These parameters are then combined
with observed image intensities to estimate the motion parameters of the object.
A large training set is created using artifical pertubations to images. The
proposed method is then tested on 7 video sequence each containing 100 to 400
frames. The topic of the paper is very relevant to the workshop.

The approach used is clearly explained in the paper. The authors proposed
parameters sensitive LLiP (PLLiP) is shown to perform better than LLiP alone.
The proposed algorithm seems to be a nice way to fuse motion and appearance
features to create a combined model for object tracking.

While the authors did not show that convergence was guaranteed, they did demonstrate
that it converged experimental using the available test data.

In the description of the training method it is not stated whether the training
examples are taken from the same sequences as the test data. If they are from the
same sequences then the generalization performance of the approach is not really
being tested.

Another general reservation I have is the amount of test data used. Seven sequences
of 100 to 400 frames gives a maximum of 2800 frames. While I realise labeling test
data is expense, perhaps the algorithm could have been run on some longer sequences
and the results qualitatively assessed. This is particularly true in Section 5.2
where the relationship between the appearance parameters and actually object appearance
is demonstrated using one sequence, with one head turn.

In Figure 1 there is a problem with the format, the text should not run beside the
figure. You should confirm you are using the format style files provided.
In Table 1 the label HEAD 2 is repeated, perhaps this should be HEAD 3.


-------------------- review 2 --------------------

PAPER: 28
TITLE: Simultaneous learning of motion and appearance

OVERALL RATING: 2 (accept)
REVIEWER'S CONFIDENCE: 2 (medium)
----------------------- REVIEW --------------------

This paper introduces a new learning method for motion estimation of objects.
The feature is the appearance mapping and motion estimation method can be optimized
simultaneously and appearance parameters are estimated by unsupervised learning.
The experiments show the method can deal with strong variable illumination, non-rigid
deformations and self-occlusions.

Some equations should be more clearer, for some notations are not explicitly given