Speaker: Prof. David Suter, University of Adelaide (Australia), cs.adelaide.edu.au/~dsuter/
Location: ICG Seminar Room
Abstract: For at least 25 years, researchers in the more geometrical aspects of computer vision (3D reconstruction, segmentation based on geometrical models, etc.) have recognised that some form of robust statistical procedure is required to deal with outliers (badly corrupted measurements, model failure) and pseudo-outliers (multi-structural data). Early solutions: Ransac and the Hough Transform are still used - more or less without major change - more than 35 years after their invention. This is despite many criticisms one may make of these methods and many improvements that have been proposed to address some of these shortcomings. This talk will outline some of these shortcomings and some of the solutions (focussing mostly on work carried out by the speaker and his research associates) with a view to why these works (and the works of others) have not supplanted Ransac (in particular) and the Hough Transform to the degree one might expect.
Speaker: Prof. Philippos Mordohai, Stevens Institute of Technology (US)
Location: HS i11
Abstract: After several years of machine learning being a driving force behind progress in computer vision research, machine learning methods have recently found great success in 3D reconstruction from images. In this talk, I will present how our supervised learning approach was able to improve several aspects of 3D reconstruction by leveraging data with ground truth that have recently become available in sufficient amounts.
By training classifiers on these data, we can eliminate suboptimal heuristic choices, which were a key characteristic of conventional methods, and gain the capability to predict whether two pixels correspond, whether a reconstructed 3D point is reliable, or whether a voxel in an occupancy grid is occupied. Our approach is not entirely data-driven, but allows the integration of prior knowledge in the form of constraints that have been proven effective in the long term. This is accomplished via the use of random forest classifiers that are well suited for data with heterogeneous feature vectors and by considering long-range relationships that go beyond image patches, such as left-right consistency. We present results in (i) binocular stereo matching, (ii) fusing multiple disparity maps generated by black-box matching algorithms, and (iii) in multi-view occupancy grid estimation.
Moreover, we show evidence that our approach generalizes well across datasets by training it on indoor and testing it on outdoor imagery and vice versa.
I will also briefly overview other ongoing research projects on active vision and robotic perception. Bio: Philippos Mordohai is an Associate Professor of Computer Science at the Stevens Institute of Technology. Prior to that, he was a postdoctoral researcher at the University of North Carolina, Chapel Hill, and the University of Pennsylvania. He holds the Diploma in Electrical and Computer Engineering from the Aristotle University of Thessaloniki, Greece, and the MS and PhD degrees, both in Electrical Engineering, from the University of Southern California. His research interest include 3D reconstruction from images and video, range data analysis, perceptual organization and active vision. He serves as an Associate Editor for the Journal of Image and Vision Computing, as a guest editor for the Computer Vision and Image Understanding journal, and as an area chair of WACV in 2014 and ICPR in 2014 and 2016. His research has been funded by the US National Science Foundation, National Institutes of Health, Department of Homeland Security and Google Inc. More information is available at www.cs.stevens.edu/~mordohai
Speaker: Philip Voglreiter, ICG
Location: ICG Seminar Room
Abstract: Methods for Minimally Invasive Canter Treatment present an important alternative for traditional surgical resection of tumors. However, predicting the outcome of such techniques is difficult, even for experienced interventional radiologists. Consecutively, accurate simulation can help to decide whether tumors are safely destroyed in these interventions. This is not only crucial for affected patients, but can also help cutting potential costs for necessary follow-up treatment.
In this talk, I will provide an overview of the research conducted in three EU projects, involving a total of five technical institutions and five medical universities. While two of these projects (Imppact and GoSmart) successfully finished already, the third (ClinicImppact) is still running until June 2017. The timeline of the interdisciplinary work consists of many achievements. Amongst others, our consortium achieved early fundamental results in biomechanical properties and optimized medical routines for improved accuracy of simulating treatment. However, this talk will focus on the technical achievements. I will present a number of image processing, visualization and computational simulation techniques which culminate into technical environments for simulating minimally invasive cancer treatment. One of these environments focuses on computational performance. We achieve high accuracy simulation which is up to an order of magnitude faster than real-time treatment. This indicates potential usefulness as a decision-making tool for treatment planning in the clinical routine. The second application aims at comparing different treatment methods, devices and protocols. In a web-based environment, we provide a testbed for medical experts, researchers from biomechanical simulation and device vendors. Contributors can easily implement and evaluate novel techniques for simulation, protocols for treatment or algorithms implemented in devices and benchmark them with established techniques or results of real treatment. By sharing the techniques among users, interventional radiologists can simulate treatment with a multitude of devices and protocols, and pick the optimal method for an individual patient.
Speaker: Piotr Didyk, Max Planck Institute for Informatics & Saarland University
Location: ICG Seminar Room
Abstract: There has been a tremendous increase in quality and number of new
output devices, such as stereo and automultiscopic screens, portable
and wearable displays, and 3D printers. Some of them have already
entered the mass production and gained a lot of users’ attention;
others will follow this trend promptly. Unfortunately, abilities of
these emerging technologies outperform capabilities of methods and
tools for creating content. Also, the current level of understanding
of how these new technologies influence user experience is
insufficient to fully exploit their advantages. In this talk, I will
demonstrate that careful combinations of new hardware, computation,
and models of human perception can lead to solutions that provide a
significant increase in perceived quality. More precisely, I will show
how careful rendering of frames can improve spatial resolution beyond
physical capabilities of display devices. Next, I will discuss
techniques for overcoming limitations of current 3D displays. In the
context of 3D printing, I will present methods for specifying objects
for multi-material 3D printing. Short Bio: Piotr Didyk is an Independent Research Group Leader at the Cluster of
Excellence on ”Multimodal Computing and Interaction” at the Saarland
University (Germany), where he is leading a group on Perception,
Display, and Fabrication. He is also appointed as a Senior Researcher
at the Max Planck Institute for Informatics. Prior to this, he spent
two years as a postdoctoral associate at Massachusetts Institute of
Technology (MIT). In 2012, he obtained his Ph.D. from the Max Planck
Institute for Informatics and the Saarland University for his work on
perceptual display. During his studies, he was also a visiting student
at MIT. His research interests include human perception, new display
technologies, image and video processing, and computational
fabrication. He focuses on techniques that account for properties of
the human sensory system and human interaction to improve perceived
quality of the final images, videos, and 3D prints. More info:
https://people.mpi-inf.mpg.de/~pdidyk/.
Speaker: Alexander Jung, Assistant Professor, Aalto University
Location: ICG Seminar Room
Abstract:In this talk, I present some of our most recent work on applying tools from compressed sensing to (semi-supervised) machine learning with massive network-structured datasets, i.e., big data over networks. We expect the use of compressed sensing ideas game changing as it was for digital signal processing. In particular, I will present a sparse label propagation algorithm which efficiently learns the labels for data points based on the availability of a few labeled training data points. This algorithm is inspired by compressed sensing methods and allows for a simple sufficient condition on the network structure and available label information such that accurate learning is possible.
Speaker: Alexander Shekhovtsov, Christian Reinbacher
Location: ICG Seminar Room
Abstract: Dense image matching is a fundamental low-level problem in Computer Vision, which has received tremendous attention from both discrete and continuous optimization communities. The goal of this paper is to combine the advantages of discrete and continuous optimization in a coherent framework. We devise a model based on energy minimization, to be optimized by both discrete and continuous algorithms in a consistent way. In the discrete setting, we propose a novel optimization algorithm that can be massively parallelized. In the continuous setting we tackle the problem of non-convex regularizers by a formulation based on differences of convex functions. The resulting hybrid discrete-continuous algorithm can be efficiently accelerated by modern GPUs and we demonstrate its real-time performance for the applications of dense stereo matching and optical flow.
Tuesday 19. January 2016, 13:00
Title: BaCoN: Building a Classifier from only N Samples
Speaker: Georg Waltner
Location: ICG Seminar Room
Abstract: We propose a model able to learn new object classes with a very limited amount of training samples (i.e. 1 to 5), while requiring near zero run-time cost for learning new object classes. After extracting Convolutional Neural Network (CNN) features, we discriminatively learn embeddings to separate the classes in feature space. The proposed method is especially useful for applications such as dish or logo recognition, where users typically add object classes comprising a wide variety of representations. Another benefit of our method is the low demand for computing power and memory, making it applicable for object classification on embedded devices. We demonstrate on the Food-101 dataset that even one single training example is sufficient to recognize new object classes and considerably improve results over the probabilistic Nearest Class Means (NCM) formulation.
Speaker: Georg Waltner
Location: ICG Seminar Room
Abstract: We propose a model able to learn new object classes with a very limited amount of training samples (i.e. 1 to 5), while requiring near zero run-time cost for learning new object classes. After extracting Convolutional Neural Network (CNN) features, we discriminatively learn embeddings to separate the classes in feature space. The proposed method is especially useful for applications such as dish or logo recognition, where users typically add object classes comprising a wide variety of representations. Another benefit of our method is the low demand for computing power and memory, making it applicable for object classification on embedded devices. We demonstrate on the Food-101 dataset that even one single training example is sufficient to recognize new object classes and considerably improve results over the probabilistic Nearest Class Means (NCM) formulation.