|24 November 2017, MI 03.13.010|
Invited Talk by Prof. Michel Defrise
This talk introduces the concept and applications of consistency conditions in inverse problems with redundant data. We consider inverse problems modelled by some linear operator or matrix A, such as the Radon transform in tomography. The inverse problem consists in solving an equation Ax=y for x given measured data y. In many applications this equation admits a solution only if the data satisfy a set of equations denoted C(y)=0, referred to as the consistency conditions. If the data are noise free and the operator A accurately models the imaging system, the data are by definition consistent because the “exact” object x satisfies Ax=y, and in that case C(y)=0. In practice however the consistency conditions are not satisfied; they can then be used to estimate some vector of parameters p of the imaging system (typically calibration parameters) by solving C(p, y)=0 for p, where C(y, p) denotes the consistency condition corresponding to the parameter p. After a general introduction to the concept, we will review a variety of examples pertaining to 2D and 3D tomography. Applications will be briefly described.
|24 November 2017, MI 01.07.014|
PhD defense by Matthias Wieczorek
Modern X-ray based imaging enables recording of phase-contrast (refraction) and dark-field (scattering) information. Tomographic Reconstruction of the dark-field signal poses an especially challenging problem, as the scattering within an object depends on its orientation. Within this thesis an abstract software framework for Tomographic Reconstruction as well as a novel method for Anisotropic X-ray Dark-field Tomography will be presented. A first biomedical experiment on a sample of a human cerebellum indicates that this method could provide a complementary imaging modality for imaging nerve fibers.
|22 November 2017, MI 01.09.014|
PhD defense by Kanishka Sharma
In Autosomal Dominant Polycystic Kidney Disease (ADPKD), automated segmentation of kidneys for total kidney volume (TKV) measurement is very challenging due to severe, disease-related alterations in kidney morphology. This PhD? thesis analyzes the applicability and performance of machine learning techniques (Random Forests and Deep Learning) for kidney segmentation in ADPKD. The developed segmentation method based on Deep Learning allows fast and reproducible TKV measurements, demonstrating that machine learning can be successfully used for complex medical image segmentation tasks.
|22 November 2017, MI 00.05.035|
PhD defense by Shadi Albarqouni
Aimed at improving machine learning algorithms by incorporating domain-specific knowledge, we develop a set of mathematical and technical methods that cope with different conditions of data abundance, reliable labels, and class balance. Proposed methods are evaluated for various biomedical applications, in particular, Tomographic Reconstruction and Noise Reduction in Cryo-Electron Tomography, Mitotic figure Detection in Breast Cancer Histology Images, and Depth Perception in Interventional Imaging.
|19 October 2017, FMI 00.12.019|
Invited Talk by Events.PerceptiveIO: The Need for Speed in Computer Vision
Depth cameras are becoming key tools for computer vision tasks ranging from hand, body or object tracking, 3D reconstruction and simultaneous localization and mapping. Almost all these tasks need to solve a tracking problem i.e. each new frame of depth and image data is correlated to the previous, and this temporal information allows for faithful pose and/or geometry reconstruction over time. However, this reliance on temporal information, makes tracking problems hard to solve when using sensors running at 30fps, due to susceptibility to high frame-to-frame scene motions and artifacts such as motion blur. [..] In this talk we are going to demonstrate a 3D capture system for high speed and high quality depth estimation, and show its advantages in a variety of computer vision tasks. Our hardware and software depth pipeline can run at 1.1msec with modern GPUs and readily procurable camera and illumination components.
|17 October 2017, FMI 00.12.019|
Invited Talk by Dr. Hamid Rezatofighi from University of Adelaide, Australia
Recently, deep structured networks such as deep convolutional (CNN) and recurrent (RNN) neural networks have enjoyed great success. However, similar to most machine learning techniques, current deep learning approaches rely on conventional statistics and restricted to a specific problem formulation. In particular, they are designed to learn a model for a distribution (or a function) that maps a structured input, typically a vector, matrix, or tensor, to a structured output. Due to this limitation, deep learning has not had any success or significant impact in many other applications. In fact, many problems such as object detection, graph matching, and multi-target tracking, are naturally expressed with sets of elements rather than vectors. As opposed to a vector, the size of a set is not fixed in advance, and it is invariant to the ordering of its elements. Therefore, learning approaches built based on conventional statistics cannot be used directly for these problems. In the talk, I will discuss deep learning approaches that use finite set statistics and point processes instead of conventional statistics, as a step towards the set learning problem. I will also highlight further recent work.
|12 October 2017, FMI 00.12.019|
Invited Talk by Dr. Del Bue: Tomasi and Kanade revisited: from 3D point reconstruction to object based structure from motion.
This talk will present how semantic information as extracted by object and region detectors can be used to perform classical 3D Computer Vision tasks in a robust way. In particular, object positions can be used to simply calibrate cameras and to identify the position and orientation of objects in 3D. Strikingly, this solution has a closed form that can be solved very efficiently, similarly to the standard Tomasi and Kanade factorization problem. Moreover, the information of the position of objects can be used to initialise standard methods for 3D reconstruction using feature point matches. This new approach provides a novel pipeline to efficiently inject semantic information in Structure from Motion problems.
|12 October 2017, FMI 00.12.019|
Invited Talk from Prof. Cucchiara: Beyond People Detection: GANs and iterative networks working on human data
Computer Vision with CNNs performs well for people detection. This is not enough. A step forward can be taken to understand the aspect of people detected in low resolution, or corrupted by occlusions in the crowd; to track them in the wild; to detect saliency and pay attention to details only; to forecast motion and human actions. [..] The talk will present projects and results carried out at Imagelab University of Modena and Reggio Emilia on people aspect classification, tracking and saliency analysis based on what people see.
|6 October 2017, |
PhD Defense by Oliver Zettinig
Ultrasound imaging is commonplace in clinical routine and has become the standard of care for a plethora of diagnostic scenarios. Due to issues such as limited image quality or obstructed visibility of anatomy, the exclusive use of this modality for interventional guidance purposes has, however, not yet reached a comparable level of maturity. This thesis addresses several challenges associated with ultrasound imaging by proposing advanced techniques for interventional use. Although applied to a broad spectrum of clinical fields and anatomies, their underlying methodology is generic and can be transferred to other medical scenarios. First, a framework for multi-modal prostate biopsy guidance is introduced, allowing urologists to accurately target suspicious lesions by combining trans-rectal ultrasound information with complementary functional tomographic data. The crucial part of this fusion consisting of deformable image registration is solved by two novel algorithms based on automatically segmented prostate surfaces or a preconditioned intensity similarity metric and a statistical deformation model. To overcome the challenges of manual acquisitions, in particular navigation to and maintenance of appropriate location and suitable acoustic window, robotic solutions are studied. Based on multi-modal image registration, a visual servoing control scheme for neurosurgical navigation is introduced. While compensating for target anatomy movements in real-time, it allows for automatic needle guide alignment for accurate manual insertions. The suitability of such systems for reliable robotic acquisitions even in absence of planning data is demonstrated by applying the developed methods, including image quality optimizations using confidence maps, for automated abdominal aortic aneurysm screenings. Through Doppler modes, ultrasound physics uniquely allows fast analysis of blood flow dynamics, albeit limited to 2D projections. This thesis introduces a novel technique to recover 3D velocity information in combination with a temporal flow profile using measurements from multiple directions. Due to the importance of accurate and linearly independent sampling, the advantages of robotic acquisition schemes can be hereby fully exploited. Results of phantom experiments, volunteer studies and clinical patient evaluations, all in close collaboration with medical partners, demonstrate the great potential benefit of advanced ultrasound imaging techniques in interventional settings in terms of both efficacy and efficiency.
|2 October 2017, |
NeuBtracker: publication in Nature Methods
The work on the open-source microscope NeuBtracker by Prof. Dr. Gil Westmeyer and his group was carried out with participation from CAMP. The results have now been published in Nature Methods, see also the TUM press release.
|8 August 2017, |
One paper accepted at IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
|5 August 2017, Vancouver, Canada|
2 papers accepted at IROS 2017
|25 July 2017, Venice, Italy|
5 papers accepted at ICCV 2017
We are happy to announce that 5 papers of our chair have been accepted at ICCV
|25 July 2017, Quebec City, Quebec, Canada|
11 papers accepted at MICCAI 2017
We are happy to announce that 11 papers of our chair have been accepted at MICCAI