问题描述:
英语翻译
In this paper,we substantialize the general adaptation
problem as a subspace adaptation problem in nonstationary
appearance tracking,where the target visual appearances fora short time interval are represented as a linear subspace.We
analyze the ill-posed adaptive tracking problem in this setting.
In our approach,we enforce three novel constraints for the
optimal adaptation:1) negative data that are easily available,
2) pair-wise data constraints that are used to identify positive
data from bottom-up,and 3) adaptation dynamics that smooth
the updating process.Then,we give a closed-form solution to
this subspace tracking problem and also provide a practical
iterative algorithm.Our method not only estimates the motion
parameters of the target but also keeps track of the appearance
subspaces.Note our model adaptation strategy is fundamentally
different from the methods using on-line learning [23],
[25],where the appearance models at previous time frames are
directly utilized to supervise labelling the current image observations,
and to determine the new information for updating
the model.In contrast,we mainly resort to the data-driven
constraints to select positive and negative data from current
image observations for the appearance model adaptation.
In the next section,we briefly review related tracking algorithms
in terms of different observation models,afterwards the
dilemma in the traditional adaptation schemes is investigated.
Our solution to the dilemma of the adaptation is elaborated in
Section III.In Section IV,we present the experiments results
and discussions.The concluding remarks are given in Section V.
II.RELATED WORK AND MOTIVATION
In tracking,the target is tracked or detected based on the
matching between the observed visual evidence (or measurements)
and the visual model.We generally call the model that
stipulates the matching as the observation model or the measurement
model.Without any prior of the targets,appearancebased
tracking approaches are more general than feature-based
methods.The visual appearances of an object may bear a manifold
in the high-dimensional image space.Depending on the
features used to describe the target and on the variances of the
appearances,such a manifold can be quite nonlinear and complex.
Therefore,the complexity in the appearances largely determines
the degrees of difficulty of the tracking task.
In this paper,we substantialize the general adaptation
problem as a subspace adaptation problem in nonstationary
appearance tracking,where the target visual appearances fora short time interval are represented as a linear subspace.We
analyze the ill-posed adaptive tracking problem in this setting.
In our approach,we enforce three novel constraints for the
optimal adaptation:1) negative data that are easily available,
2) pair-wise data constraints that are used to identify positive
data from bottom-up,and 3) adaptation dynamics that smooth
the updating process.Then,we give a closed-form solution to
this subspace tracking problem and also provide a practical
iterative algorithm.Our method not only estimates the motion
parameters of the target but also keeps track of the appearance
subspaces.Note our model adaptation strategy is fundamentally
different from the methods using on-line learning [23],
[25],where the appearance models at previous time frames are
directly utilized to supervise labelling the current image observations,
and to determine the new information for updating
the model.In contrast,we mainly resort to the data-driven
constraints to select positive and negative data from current
image observations for the appearance model adaptation.
In the next section,we briefly review related tracking algorithms
in terms of different observation models,afterwards the
dilemma in the traditional adaptation schemes is investigated.
Our solution to the dilemma of the adaptation is elaborated in
Section III.In Section IV,we present the experiments results
and discussions.The concluding remarks are given in Section V.
II.RELATED WORK AND MOTIVATION
In tracking,the target is tracked or detected based on the
matching between the observed visual evidence (or measurements)
and the visual model.We generally call the model that
stipulates the matching as the observation model or the measurement
model.Without any prior of the targets,appearancebased
tracking approaches are more general than feature-based
methods.The visual appearances of an object may bear a manifold
in the high-dimensional image space.Depending on the
features used to describe the target and on the variances of the
appearances,such a manifold can be quite nonlinear and complex.
Therefore,the complexity in the appearances largely determines
the degrees of difficulty of the tracking task.
问题解答:
我来补答展开全文阅读