Allow young researchers (primarily MSc/BSc students) to present their ongoing work.

This paper gives a short summary of work related to the creation of a department-hosted hand vein database. After the introducing section, special properties of the hand vein acquisition are explained, followed by a comparison table, which shows key differences to existing well-known hand vein databases. At the end, the ROI extraction process is described and sample images and ROIs are presented.

In this work we show that the classification performance of high-dimensional structural MRI data with only a small set of training examples is improved by the usage of dimension reduction methods. We assessed two different dimension reduction variants: feature selection by ANOVA F-test and feature transformation by PCA. On the reduced datasets, we applied common learning algorithms using 5-fold cross-validation. Training, tuning of the hyperparameters, as well as the performance evaluation of the classifiers was conducted using two different performance measures: Accuracy, and Receiver Operating Characteristic curve(AUC). Our hypothesis is supported by experimental results.

Featured talks offer an opportunity to present outstanding work (already accepted to top-ranked conferences/journals in the field) to the community.

In this work we consider the problem of solving multiple prediction tasks. Though traditional machine learning techniques can be applied to solve each task in isolation, they usually require extensive amounts of labeled data in order to archive reasonable generalization accuracy. In contrast, multi-task algorithms solve the tasks jointly by sharing information between them. Both experimental and theoretical results have shown that it allows better generalization from fewer training points per task. However, the success of information transfer highly depends on the level of relatedness between the tasks. In particular, in realistic scenarios not all the tasks might be equally related and hence it could be preferable to transfer only between the most related tasks. We propose an approach that processes tasks in a sequence and allows information transfer only between the subsequent tasks instead of solving all of them simultaneously. This makes learning more flexible in terms of variability between the tasks and also memory efficient as it does not require processing all training samples at the same time. Subsequently, we address the problem of curriculum learning of tasks, namely, the role of the order in which tasks are being solved in terms of the average accuracy. We study this question using PAC-Bayesian theory and prove a generalization bound that quantifies the effectiveness of the order and therefore can be used for choosing a beneficial order. Based on this result we propose a theoretically justified algorithm that chooses a favorable sequence for learning by minimizing the value of the generalization bound in the greedy manner. Our experimental results on two image datasets show that learning tasks in a sequence can be more effective than learning each task independently or all of them jointly. They also show that the order in which tasks are being solved can substantially influence the overall performance and that the proposed algorithm is able to automatically discover a beneficial task order. This work was accepted for publication at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

The topic of crowd modeling in computer vision assumes a single generic typology of crowd, which is very simplistic. In this paper we adopt a widely accepted taxonomy for crowds, focusing on a particular category, the spectator crowd, which is formed by people interested in watching something specic that they came to see" 1. This can be found at the stadiums, amphitheaters, cinema, etc. In particular, we propose a novel dataset, the Spectators Hockey (S-Hock), which deals with hockey matches during an international tournament. The dataset considers 4 hockey matches, where hundreds of spectators are individually annotated, capturing ne grained actions such as hands on hips, clapping hands, watching the cellphone etc., for a total of more than 100 millions of annotations. Analyzing people at the stadium addresses dierent computer vision tasks, some of them are classic (crowd counting), while other are brand new (as the spectator categorization). For this reason, S-Hock comes also with a set of protocols for dealing with all of them, and a set of baselines and novel approaches that dene the best scores on all the tasks. Anyway, the performances are far from being errorless, and this witnesses the diculty of the problem and that much can be done in the future. This work has been accepted as poster presentation at CVPR 2015 in Boston.

Visual object tracking has been an active area in computer vision for decades as many open scientific questions still remain. Parts-based models might have advantages to holistic representations as the modelling of local information seems advantageous for the problem of re-identification after full occlusion. This talk introduces a novel method for establishing keypoint correspondences on objects for single-target object tracking. The key ingredients are keypoint redundancy as well as a dissimilarity measure between correspondences that takes into account their geometric compatibility and consensus, allowing the approach to separate inlier correspondences from outliers. We employ both static correspondences from the initial appearance of the object as well as adaptive correspondences from the previous frame to address the stability-plasticity dilemma. The geometric dissimilarity measure enables also to disambiguate keypoints that are difficult to match. Based on these ideas we build a keypoint-based tracker that outputs rotated bounding boxes. We demonstrate in a rigorous empirical analysis that this tracker outperforms the state of the art on a dataset of 77 sequences. This work was accepted for publication at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

This work presents a flexible system to reconstruct 3D models of objects captured with an RGB-D sensor. A major advantage of the method is that unlike other modelling tools, our reconstruction pipeline allows the user to acquire a full 3D model of the object. This is achieved by acquiring several partial 3D models in different sessions that are automatically merged together to reconstruct a full model. In addition, the 3D models acquired by our system can be directly used by state-of-the-art object instance recognition and object tracking modules, providing object-perception capabilities for different applications, such as human-object interaction analysis or robot grasping. The system does not impose constraints in the appearance of objects (textured, untextured) nor in the modelling setup (moving camera with static object or a turn-table setup). The proposed reconstruction system has been used to model a large number of objects resulting in metrically accurate and visually appealing 3D models.

Reliable and exact assessment of visibility is essential for safe air traffic. In order to overcome the drawbacks of the currently subjective reports from human observers, we present an approach to automatically derive visibility measures by means of image processing. It first exploits image based estimation of the atmospheric transmission describing the portion of the light that is not scattered by atmospheric phenomena (e.g., haze, fog, smoke) and reaches the camera. Once the atmospheric transmission is estimated, a 3D representation of the surrounding (digital surface model: DMS) is used to compute depth measurements for the haze-free pixels and then derive a global visibility estimation for the airport. Results on foggy images demonstrate the validity of the proposed method.

An important focus of current research in the field of Micro Aerial Vehicles (MAVs) is to increase the safety of their operation in general unstructured environments. Especially indoors, where GPS cannot be used for localization, reliable algorithms for localization and mapping of the environment are necessary in order to keep an MAV airborne safely. In this paper, we compare vision-based real-time capable methods for localization and mapping and point out their strengths and weaknesses. Additionally, we describe algorithms for state estimation, control and navigation, which use the localization and mapping results of our vision-based algorithms as input.

In computer vision, we have often the problem of creating a graph out of an unstructured point-set, the so-called data graph. A common solution for this problem is to build a Delaunay triangulation, which is the dual of the Voronoi diagram. Although popular, there has been little research on alternatives to Delaunay triangulation which could be tuned towards specific applications such as point-set registration. A good research direction could focus on the stability of the data graph, especially when points suffer variations due to noise. After building a data-graph, one could perform matching between two graphs in order to register the original point-sets. In this paper, we propose a data graph technique based on the Minimum Spanning Tree of Maximum Entropy (MSTME). We are aiming at a data graph construction which is more stable with respect to small variations in the neighborhood of points when compared with Delaunay, but also, we are focusing on a graph which could help the point-set registration process. We propose an algorithm with a single free parameter that weights the importance between the total weight cost and the entropy of the current solution. We compared our algorithm on a number of different databases with the Delaunay triangulation.

In multimedia research, we often focus on compression, transmission, storage, analysis and retrieval of data. In human computer interaction, we investigate how users operate computers and access the computing capabilities. However, sometimes the actual why is not subject to research. The goals and aims of users, i.e., their intentions, have a strong influence on what people do with data, especially images and videos. The relation between users intentions and data, however, has often been recognized and sometimes even named (intention gap, query gap). This talk investigates and summarizes research on the relation between why, the goals of users, and how, the support multimedia information and retrieval systems can provide.

Present recently funded, or already running, projects to the community in order to encourage potential collaborations in the future.

Gallium, indium, and certain rare-earth elements are mandatory for the production of modern electronics such as Integrated Circuits (ICs), photovoltaics, and flat panel displays. The production of these elements is, however, geographically concentrated, and their demand has been rising faster than the supply. The goal of the Reclaim project (re-claim.eu) is to overcome this problem by developing technologies for reclaiming these chemical elements from electronics waste. We present our ongoing research for this project, which focuses on analyzing images and videos of Printed Circuit Boards (PCBs) for recycling purposes. We discuss suitable image acquisition setups and present two methods for PCB analysis. The first method allows for monitoring waste streams for certain (e.g. valuable or hazardous) PCBs in real-time to enable automatic sorting. This is accomplished via local feature matching to ensure invariance with respect to the position and orientation of PCBs in the waste stream as well as robustness to dust and PCB damage. The purpose of the second method is to count and localize IC chips in high-resolution images for automatic component disconnection. We show that this task can be achieved using state-of-the-art object detection methods and investigate means for improving the performance by incorporating priors on the size and arrangement of IC chips. The performance of both methods is assessed on extensive public datasets

Acute Lymphoblastic Leukemia (ALL) is the most frequent leukemia entity in children and adolescents. Despite continued progress and refinement of therapeutic approaches, disease relapse due to insufficient extinction of leukemic blasts still remains the number one cause of treatment failure. About 15-20% of pediatric patients with the disease still suffer from relapse. Flow cytometry (FCM) is one of the methodologies most useful in this respect, because it is widely available and applicable to most patients. While FCM acquisition can be harmonized straightforwardly, data analysis and interpretation rely largely on operator skills and experience. Hence, these requirements represent the current bottleneck of safely applying the FCM-Minimal Residual Disease (MRD) methodology in a growing community of diagnostic laboratories. Thus, Machine Learning (ML) methods represent an objective methodology for automated analysis of FCM data to cluster/classify cell populations and determine MRD. Although automatic clustering methods based on mixture models are suitable for general FCM data there are currently no comprehensive evaluations, particularly for ALL data. Therefore, we concentrate on the evaluation of different strategies for automatic FCM analysis on a newly collected database: the annotated data set derives from samples of ALL patients that have been treated uniformly according the AIEOP-BFM 2009 treatment protocol. In detail, the data set will initially hold about 200 leukemia cases for B cell precursor ALL and will be expanded to T-lineage ALL. Representatives of all known sample appearances including phenotypic shifts, subclone formation, and parallel occurrence of normal regenerating precursor cells are present in the dataset. Many samples of the 200 cases hold low MRD-rates (<1%), since reaching state-of-the-art detection accuracy below this proportion is crucial. Each leukemia case is measured at 3 time points during leukemia treatment (d15, d33 and d78), with 2 tubes per time point, and each containing 300.000 cells with currently 9 FCM-parameters.All samples have been annotated using a standard gating procedure that defines the operator based ground truth of blast cells. The dataset will be extended by ALL data of the same leukemia types from different medical centers and different FCM devices, to develop a device independent methodology. Thus, the data will be normalized to be independent of different channels and setup variations. The data set is used for an evaluation of different learning approaches (Mean-Shift clustering with a Bayes decision, Bayes decision based on Kernel Density Estimation (KDE), and a Bayes Classifier based on mixture models). The usage of a Gaussian Mixture Model allows a compact representation of the high dimensional FCS data so we propose a model based on GMM, learned on a training set of 150 patients. Blasts (leukemic) cells are modeled separately from the non-blasts in order to prevent that the first group is missed in the modeling phase. The model is then represented using a set of basis functions in order to reduce the computational complexity in the testing phase. Preliminary results show that automatically classified populations deviate from the ground truth by an error range that is within the inter operator variance. A summary of the evaluation on our ALL data set is presented and future perspectives are discussed.

In this talk, the 'Centre of Image and Material Analysis in Cutural Heritage' (CIMA) is introduced. The centre is devoted to the investigation of ancient manuscripts and is part of an interdisciplinary cooperation between philologists, chemists and computer scientists. The historical manuscripts investigated by CIMA so far originate from the Middle Ages and are written in different languages and scripts and are often in a poor condition. The transcription of the historical texts is one of the main parts of the work conducted by the philologists. However, such a transcription is impeded by the circumstance that the documents are often in a bad condition, e.g. they contain erased or faded-out text or are corrupted by background clutter. MultiSpectral Imaging (MSI) has proven to be a valuable tool for the investigation of such objects, since it is capable of increasing the contrast of faded-out characters, compared to ordinary white light. Therefore, the computer scientists of CIMA apply MSI in order to increase the legibility of the texts. The MSI system used is comprised of two cameras, narrowband illumination and optical filters. While the imaging in narrow-band spectral ranges already gains a legibility increase, sev- eral post-processing techniques have been applied to the multispectral images in order to enhance the contrast of the writings furthermore. The post-processing techniques applied are based on digital image processing techniques - like local contrast enhancement or noise removal and on dimension reduction techniques. These post-processing techniques have been partially judged as useful by the philologists for the transcription of the historical texts. The presentation will con- tain several representative resulting images, which exhibit the usefulness of such post-processing methods. The philologists belonging to our team are not only transcribing the ancient texts, but are also identifying scribes of such ancient writings. Currently, paleographers are performing this task mainly manually in order to localize, date or authenticate historical writings. In order to support this time intensive work, we have recently developed an algorithm for the automated identification of writers of such documents. The method is based on local features, namely the Scale Invariant Feature Transform, which are used for the generation of a visual vocabulary. The algorithm has been tested on five historical manuscripts, which have been written by seven different scribes. Numerical results will be presented, which show that the method is capable of correctly identifying the scribes of the writings in the majority of the cases considered.

The latest estimate (Apr 17th , 2015) of the Counterfeit Goods Market Value is $653.77 Billion. Besides financial loss, counterfeit products put the brand’s reputation and in the worst case the consumer’s health at stake. Authentic Vision has developed the UniqueLabel comprising a machine-readable code and a diffractive foil (refer figure 1). The UniqueLabel is practically irreproducible and can be attached to any kind of product. It allows consumers to verify the product’s authenticity - thus enabling them to protect their and the brand’s interests in authentic products. This is possible by scanning the label with standard devices, e.g. a smartphone. The patented authentication system operates by checking the uniqueness of a label, which is basically validating its registered random optical features. Among others, systemic randomness in the production of the UniqueLabel can be used as a feature here. In our talk we give a brief introduction to our system, over-viewing also non-authentication components adding value for consumers and brand owners. We’ll discuss a simple model of the two-step printing process used to manufacture the UniqueLabel and identify the sources and quantities of the systemic randomness involved. Finally, we’re shedding some light on the feature extraction and validation of such randomness and how it can be used in the authentication process. We focus on already mastered and challenges ahead when using mobile devices in uncontrolled environments for such authentication tasks.

The 3D documentation of the tunnel surface during construction requires fast and robust measurement systems. In the solution proposed in this paper a single camera is set-up during tunnel advance in several positions, taking pictures of the tunnel surface. The captured images are assembled to a 3D tunnel surface model in a fully automatic processing chain. Image data capture is realized by the tunneling/advance/driving personnel close to the tunnel face (= the front end of the advance). Based on the following fully automatic analysis/evaluation, a decision on the quality of the outbreak can be made within a few minutes. This paper describes the system and conditions of image capture as well as the stereo-photogrammetry based workflow for the 3D reconstruction of a continuously merged 3D dense reconstruction of the entire advance region. Geo-reference is realized by means of signalized targets that are automatically detected in the images. We report on the results of recent testing under real construction conditions, and conclude with prospects for further development in terms of on-site performance.

In image processing, it can be a useful pre-processing step to smooth away small structures, such as noise or unimportant details, while retaining the overall structure of the image by keeping edges, which separate objects, sharp. Typically this edge-preserving smoothing process was achieved using edge-aware filters. However such filters may preserve unwanted small structures as well if they contain edges. In this work we present a novel framework for edge-preserving smoothing which separates the process into two different steps: First the image is smoothed using a blurring filter and in the second step the important edges are restored using a guided edge-aware filter. The presented method proves to deliver very good results, compared to state-of-the-art edge-preserving smoothing filters, especially at removing unwanted small structures. Furthermore it is very versatile and can easily be adapted to different fields of applications while at the same time being very fast to compute and therefore well-suited for real time applications.

Age estimation of individuals is important in human biology and has various medical and forensic applications. Recent interest in magnetic resonance (MR) based methods aims to investigate alternatives for established methods involving ionizing radiation, i.e. methods based on X-ray images. Automatic, software-based methods additionally promise improved estimation objectivity. We will show how informative automatically selected image features are regarding their ability to detect anatomical structures in three-dimensional (3D) images and discriminate age by exploring a recently proposed age estimation method for 3D MR images based on random regression forest. Data set of 102 3D MR images of left hand are used to evaluate age estimation performance, consisting of bone and epiphyseal gap volume localization, computation of one age regression model per bone mapping image features to age, and fusion of individual bone age predictions to a final age estimate. Qualitatively, we will demonstrated how feature selection works and which image features of skeletal maturation are automatically chosen to model the nonlinear regression function. Quantitative results of the automatic method shows an age estimation performance with a mean absolute difference of 0.85 years (std: 0.58 years) to chronological age as determined by a cross-validation experiment. The results are in line with radiologists performing visual age estimation based on X-ray projection images, thus involving radiation exposure.

Topological data analysis offers a rich source of valuable information to study vision problems. Yet, so far we lack a theoretically sound connection to popular kernel-based learning techniques, such as kernel SVMs or kernel PCA. In this work, we establish such a connection by designing a multi-scale kernel for persistence diagrams, a stable summary representation of topological features in data. We show that this kernel is positive definite and prove its stability with respect to the 1-Wasserstein distance. Experiments on two benchmark datasets for 3D shape classification/retrieval and texture recognition show considerable performance gains of the proposed method compared to an alternative approach that is based on the recently introduced persistence landscapes. This work was accepted for publication at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

We address the problem of model-free online object tracking based on color representations. According to the findings of recent benchmark evaluations, such trackers often tend to drift towards regions which exhibit a similar appearance compared to the object of interest. To overcome this limitation, we propose an efficient discriminative object model which allows us to identify potentially distracting regions in advance. Furthermore, we exploit this knowledge to adapt the object representation beforehand so that distractors are suppressed and the risk of drifting is significantly reduced. We evaluate our approach on recent online tracking benchmark datasets demonstrating state-of-the-art results. In particular, our approach performs favorably both in terms of accuracy and robustness compared to recent tracking algorithms. Moreover, the proposed approach allows for an efficient implementation to enable online object tracking in real-time.