Motion Analysis, Structure from Moion
Geometric Analysis of SFM
The problems I have investigated are related to basic processes in the perception of three-dimensional motion, shape and their relationship. Many psychophysical experiments as well as computational considerations convincingly show that actual systems cannot estimate exact distances and shapes, but instead derive a distorted version of space. What I have been studying is the geometric distortion between perceptual space (computed space) and actual space, which is difficult to model due to the complexity of the scene (scene can change unpredictably, and the traditional approach to this problem is to adopt a statistical approach with a lot of assumptions). The geometrical approach of our work is helpful towards obtaining an intuitive grasp of this complex problem.
From these geometric studies, many ambiguities inherent in the structure from motion problem emerge. A clear understanding of these ambiguities is in turn crucial to the development of more robust algorithms. Our modeling of depth distortion also points to the rich synergistic relationship between motion estimation and depth estimation. While the traditional approach has been to remove depth from the equations so as to formulate an elegant algorithm, this elegance is obtained at the expense of stabilities of the recovery process. Our study points to the potential offered by simultaneous estimation of motion and depth.
I also attempt to understand from these results the space-time representation reconstructed inside our head. Since metric shape cannot be computed in practice, vision systems have to compute a number of alternative space and shape representations for a hierarchy of visual tasks, ranging from obstacle avoidance through homing to object recognition. My current research interests are concerned with understanding the large spectrum of these representations. One possible representation is that of the ordinal depth; the metric aspects are ignored but only the order of the depth points are represented. The advantage is that ordinal depth can be robustly recovered. The following figure shows the ordinal depth recovered, rendered in colour (cold colors represent near depths, warm colors represent far depths). Currently, I am working on using ordinal depth representation for tasks such as landmark-based navigation (see Robust scene recognition).
Fig 1. Left: Scene in view. Right: its ordinal depth
reconstruction.
I
have continued my work in motion analysis. One main thrust of my research now
is in using sparse signal representation techniques to address the 3D
motion segmentation problem, as well as change detection in video sequences
with complications caused by complex and
dynamic behavior such as ocean waves, waving trees, rain and moving clouds.
Motion Segmentation
Motion segmentation
is a challenging problem in visual motion analysis. The idea is to segment the
scene into multiple rigid-body motions, based on the point trajectories or
optical flow observed in multiple camera views. It is a challenging problem
because it requires the simultaneous estimation of an unknown number of motion
models, without knowing which measurements correspond to which model. This
problem can be cast as a subspace clustering problem in which point
trajectories associated with each motion are to be clustered together. Recent
works [Elhamifar & Vidal, 2012, Liu 2010, Rao et
al., 2009] introduced compressed sensing techniques to subspace segmentation. We
seek to extend these sparsity-based techniques as
there are many difficulties with the current motion segmentation techniques. For
instance, most current techniques cannot automatically estimate the number of
motion clusters, and can only tolerate a small amount of missing entries and
outliers. Our research include: 1) to better capture the global structures of
data than current sparse techniques and to better estimate the number of motion
clusters via the mixed norm approach; 2) to better handle the conditions where
missing data, noise, and outliers are prevalent, and when the motions are
partially dependent.
Compared to the recent LRR formulation
proposed in [Liu 10], we wish to recover not only a low rank coefficient matrix
Z, but also a coefficient matrix that has a particular sparsity
pattern: the non-zero values appear in blocks, though not necessarily of the
same length for each block. We propose minimizing a mixed L2/ L0 norm of the
matrix Z comprising of two terms:
min. ||Z||2,0
+ ||Z||0,2 + ||E||0 s.t.
X = XZ + E
This problem can be relaxed into the
following version:
min. ||Z||2,1
+ ||Z||1,2 + ||E||1 s.t.
X = XZ + E
This mixed norm formulation encourages a
common set of trajectories chosen for self expression, and thus facilitates the
estimation of the number of motion groups. Due to the recent advances [Ding,
Sun, Toh] in computing the Moreau-Yoshida
regularizations of the mixed-norms in the above problem, we are excited by the
possibility of being able to efficiently solve the model as well as
demonstrating the model’s capability.
Change
Detection
We assume that any variation in
the background’s appearance (whether caused by intensity change, or
non-stationary background) are highly constrained. Therefore a matrix composed
of vectorized image frames can be approximated by a
low-rank matrix, and the moving foreground objects can be detected as sparse
outliers in this low-rank representation. In other words, a matrix composed of vectorized image frames D’ can be approximated as the sum
of a low-rank matrix D representing the background with its highly redundant
nature, and a sparse matrix E consisting of the moving objects treated as
sparse outliers. We then seek to solve the following optimization problem:
minD,E
rank(D) + λ ||E||0
s.t.
D’ = D + E
where E is a sparse matrix with most of its entries being zero,
and λ is an appropriate regularizing parameter.
The ability to handle target objects
of different scales is also a perennial challenge in many segmentation problems
and is also present in the preceding low-rank formulation (manifested by the
regularizing parameter λ. The root of this problem lies in that the
precise definition of the foreground target is intricately linked with the
object of interest in the scene (i.e. one’s purpose) and can be well defined
only if the object of interest or its salient characteristics is known to us.
However, knowing about the object of interest even before segmenting the scene
seems to make the problem as one of many chicken-egg problems in computer
vision, as we usually need to segment the scene in order to recognize the
objects in it. So, how can we identify an object and know its probable size
even before segmenting it? Our proposed approach is to make a weak
identification of the object of interest such as a rough region of interest.
Obtaining these weak identifications without doing any segmentation is not a
difficult problem as it can be done using low-level saliency cues in the visual
attention systems.
Resolving the scale
issue allows us to fully realize the potential of the RPCA method, without
having to resort to smoothness prior such as MRF, thus allowing targets of
various sizes to be delineated cleanly under various challenging conditions
mentioned above. The significant improvement in the accuracy and the acuity of
detection compared to the state-of-the-art techniques can be seen in the following
results
o
Non-stationary Background: this depicts a campus scene, with a large
swaying tree and a person walking along the walkway. The magnitude of the tree
motion is larger than that of the human. As can be seen from the results, the
human silhouette is cleanly delineated in the foreground mask, which is
typically not the case for many other methods that imposed smoothness prior on
the foreground shape. In the background extracted, there is no ghostly presence
resulting from residual human presence not cleanly removed. When viewed in
video sequence, the effect is much clearer, and it can also be seen that the
motion of the swaying tree is largely retained in the background.
Foreground detected
Background subtracted
o
Falling rain causing “motion” in the background. Small drizzle does not
cause any problem. We have also tested a sequence taken under a tropical
thunderstorm; the results are much poorer, and more works need to be done.
Foreground detected
Background subtracted
o
Night scene with high image noise, and flickering illumination caused by
fluorescent lights. The change in intensity caused by the varying illumination
is not a problem as it is entirely captured by the low rank constraint. We have
tested our method on other clips taken during dawn and dusk hours, and there
the rapidly changing illumination does not cause any problem too.
Foreground detected
Background subtracted
o
Boat scenes with waves motion in the background. Note that both the large
boat in the middle and the small boat in the far distance are detected,
demonstrating the scale-invariance of our method.
Foreground detected
Background subtracted