After just
over ten years with Microsoft
Research Cambridge I, Carsten Rother, will be
moving to Dresden/Germany in October 2013 to start the new Computer Vision
Lab Dresden (CVLD) as full (W3)
Professor at TU
Dresden Faculty of Computer Science. Below you will find the mission
statement and future research areas. Further Information about teaching, lab members,
and other activities in the lab will be announced at a later stage.
Looking for people! Join me in the exciting times of
starting a new lab. I have several openings for Postdocs and PhD students.
I am looking for motivated students who are excited to conduct high-quality
research in one of the research areas listed below. You should have a strong
mathematical background, ideally in optimization, machine learning, and
probability theory. You should also have good programming skills and ideally
have coded a small-scale vision application in the past. If you think that you
are the right candidate please send me your CV and letter of motivation: Professor
Carsten Rother (CVLDapplication@gmail.com).
Mission statement Computer Vision is a science that
develops models and methods for understanding, analysing, acquiring and
processing images, and more generally high-dimensional “visual” data. Computer
Vision is part of many other fields such discrete optimization, machine
learning, human computer interaction, computer graphics, and systems biology.
The Computer Vision Lab Dresden works on a broad set of topics in computer
vision and these related fields. The ultimate goal of the lab is to develop
novel theoretical concepts which are practically relevant. The emphasis is on
novelty and creativity, from both a theoretical and practical viewpoint, in
order to influence the research field and potentially set new trends. While the
set of applications is large and diverse, we often use the same modelling
language of undirected graphical models. We are interested in finding the best
trade-offs between model complexity, efficient and global optimization, and
good learning. The labs’ activity can be grouped into the following six
research areas. There are many potential synergy effects between these areas
which we try to exploit. Note that most research activities have not yet
started. For related work and previous research activities please see the Microsoft webpage.
Research Areas
Interactive Image and Data manipulation Many tasks in computer vision need the user in the loop, such interactive image segmentation, where the end result should be perceptually pleasing. We address various traditional interactive “low-level vision” topics. Additionally we plan to look at upcoming research topics such as interactive techniques for both Big Data analysis, which often occur in biological settings, as well as human-computer-interaction scenarios. Both application fields are in co-operation with partners from the respective research field. A shortlist of potential topics:
•
Interactive
Image and Video segmentation, matting, retargeting, and editing
•
Unsupervised
segmentation and co-segmentation from a single or multiple images
•
Perceptual
“loss” functions for low-level vision, such as perceptual loss for image
de-noising
•
Active
and Interactive learning in Big Data Scenarios
•
Learning
to interact, such as learning a gestural interface and activity forecasting
People: Andrew Viergutz
(PhD Candidate from Raimund Dachselt
Lab at TU Dresden).
Project webpages: to come; please visit my MSR webpage
Applied Optimization, Models, and Learning The other
research areas often require complex structured models, such as models with higher-order
functions between many variables, which pose challenges on the optimization
side. Furthermore, it is often hard to hand-code such functional relationship,
and it derive them from ground truth measurements. The goal of this theme is to
analyse the trade-offs between models, optimization and learning with the
ultimate goal of achieving an algorithm which is as efficient and accurate as
possible at test time. A shortlist of potential topics:
•
Efficient
inference for models with higher-order factors, such as object-connectivity
•
Developing
novel models such as the Decision Tree Field model, cascaded models, or Deep
Belief Nets
•
Learning
models with complex loss functions, such as Interaction Loss or Perception Loss
•
Inference
in structured models with continuous or very large label spaces, such as PMBP
•
Comparing
generative and discriminative models in real-world settings
People: Dmitrij
Schlesinger
Project webpages: to come; please visit my MSR webpage
Inverse rendering from moving images An ultimate goal in computer vision is to invert
the image formation process, ideally coming up with a graphics script that
models the geometry of the scene, camera, material and light such that the script
renders the given set of images. While this has been studied intensively in
isolation such as optical flow estimation, shape-from-shading and intrinsic
image extraction, we plan to study the problem from a holistic viewpoint in
order to optimally exploit synergy effects in a single, statically sound,
framework. A potential shortlist of topics is:
·
Dense
motion from moving cameras, such as stereo and optical flow matching
·
Image
de-convolution and de-noising.
·
Recovering
light, shading and material (including simple BRDFs)
·
Combining
statistical priors with physically motivated priors
People: Michael Hornacek (PhD
candidate, jointly supervised with Margrit Gelautz from TU Vienna),
Project webpages: to come; please visit my MSR webpage
3D Scene Understanding The ultimate goal of 3D Scene Understanding is
to take a single RGB image and extract a so-called “scene-graph”, where each
object is defined by a name, associated attributes, and its physical 3D
properties (e.g. 3D bounding box). Additionally relationships between objects
are given, such as “object A is supported by object B”. While this is hot
research area with many competing efforts, we plan to tackle the problem from a
slightly different angle. Firstly, we try to show that training models with
large amount of synthetic data is a valuable option. Secondly, we plan to
combine 3D Scene understanding with ”inverse rendering
from moving images”, in order to exploit synergy effects.
· Pose estimation and segmentation of
known objects and objects classes
· unsupervised 3D object candidate
extraction from stereo (or moving) images
· Inference with global constraints,
such as object stability
· 3D scene completion
People: Eric Brachman, Alexander Krull and Frank Michel (PhD Candidates from Stefan Gumhold’s group at TU Dresden)
Project webpages: to come; please visit my MSR webpage
Benchmarking and Label collection For
learning models as well as comparing methods it is essential to gather high
quality data and associated labels. We plan to gather new datasets and extend
existing ones. Recently, the trend of gathering synthetic data as an
alternative to real data has become popular. We are interesting in analysing
this option. Most topics in this theme are in close collaboration with other
partners. A shortlist of potential topics:
· Synthetic and real data for
low-level vision, such as matting, material estimation, etc.
· Synesthetic and real data for 3D
scene understanding
· Benchmarking Optimization techniques
for discrete undirected graphical models
Students: to come.
Project webpages: to come; please visit webpages of past projects, where
I have been involved: alpha matting
, MRF minimization, OpenGM2
Image Analysis for System
Biology Dresden
has a very strong research environment in the area of Biology. In particular
topics on Bio-Imaging grow in popularity. This is an exciting research field
for latest computer vision, optimization and machine learning ideas. The
concrete goals in this field are going to be defined soon.