Computer Vision Talk Series

Many traditional Computer Vision problems can now be solved by a single Deep Network. In this talk, I will use 3D hand tracking and interest point detection and description as examples to show that we can also tackle more complex problems by building full computer vision pipelines from multiple deep networks working together. By designing these pipelines carefully, we can keep the advantages of single Deep Networks: end-to-end learning by minimizing only one clear cost function for the full pipeline, and high performance.

The quality and robustness of CV solution directly correlates to the quality and completeness of the test data used to validate it. There is a tool to determine the quality of existing test data sets: CV-HAZOP. The talk will give a short introduction into the topic and give insights into how to apply this tool for a CV use case.

TBA.

Perception is a key component of every intelligent system as it enables actions within a changing environment. While humans perceive their environment with seemingly little efforts, computers first need to be trained for these tasks. One of the biggest challenges in computer vision are ambiguities which arise due to the complex nature of our environment and the information loss caused by observing two-dimensional projections of our three-dimensional world. In this talk, I will present several recent results in stereo estimation, 3D reconstruction and motion estimation which integrate high-level non-local prior knowledge for resolving ambiguities that can't be resolved using local assumptions alone. Furthermore, I will discuss the "curse of dataset annotation" and present a method for augmenting video sequences efficiently with semantic information.

Decomposition of graphs are a fundamental structure in discrete mathematics, with important applications in computer vision. This talk introduces the minimum cost lifted multicut problem and discusses properties of lifted multicut polytopes. It shows how applications of this problem in computer vision have advanced image segmentation, multi-object tracking and human body pose estimation.

Interactive perception leverages the capabilities of the robot to interact with the environment to reveal and understand the effects of its own actions. Knowledge about the correlations between actions and their effect is a strong prior that can be exploited to continuously interpret sensor data and enable real-time perception. We applied this insight to the estimation of the kinematic structure and state of articulated objects, and the reconstruction of their shape. Perceiving degrees of freedom is a crucial skill for robots that aim to manipulate the environment. In my talk I will present our work in interactive perception of articulated objects and other research projects from our lab that explore how interactions simplify perception and manipulation.

Self-driving cars started as a distant dream just a few decades ago. However, thanks to the recent and great progress made in many fields of computer science and engineering, this dream is now becoming reality. Self-driving cars will revolutionize our society by improving our current mobility models, and most importantly, by radically reducing the number of road fatalities. In particular, computer vision and machine learning play a central role in the perception and understanding capabilities that these vehicles required to be able to correctly operate not only under standard conditions, but also at the most unexpected situations. This talk will present the research work on road scene understanding done by the Daimler R&D Image Understanding Group during the last years, including some recent deep learning-based advancements within the context of environment perception for self-driving cars.

After introducing the classical formulation of the problem, I will focus on retrieval methods based on the bag of words image representation that exploit geometric constrains. Novel formulations of image retrieval problem will be discussed, showing that the standard ranking of images based on similarity addresses only one of possible user requirements.Retrieval methods efficiently solving the new formulations by exploiting geometric constraints will be used in different scenarios. These include online browsing of image collections, image analysis based on large collections of photographs, or 3D model construction.For online browsing, I will show queries that try to answer question such as: "What is this?" (zoom in at a detail), "Where is that?" (zoom-out to larger visual context), or "What is to the left / right of this?". For image analysis, two novel problems straddling the boundary between image retrieval and data mining are formulated: for every pixel in the query image, (i) find the database image with the maximum resolution depicting the pixel and (ii) find the frequency with which it is photographed in detail.

In this talk I will present recent research of the CG group at the HTW Dresden. In our Motion Capture Lab we have developed new methods to capture detailed surface movements of skin or apparel. With a new decomposition method (SPLOCS sparse localized deformation components) we are able to extract meaningful parameters to model, edit and control animations of highly detailed surface meshes.

The talk will firstly introduce the newly established robotics and human machine interaction lab at TU-Chemnitz. Then it will outline current research topics in robotics like human robot interaction useful in particular for flexible assembly and service-robotics. On key issue is compliant robotics, it will be explained what does compliance mean and where it is helpful and how it can be implemented as well as used for robotic applications. Moreover robot vision is very important to bring robots into our future live and into real scenarios. Therefore solution for the bin-picking and object localization problem will be shown.

Message passing solvers belong for many applications (e.g. MAP-MRF, linear assignment, b-matching, perfect matching) to the state-of-the-art, due to them being fast, having low memory footporint and being easy to implement. In the field of MAP-MRF variants of message passing (SRMP, MPLP, TRWS) have been developed in recent years, which are monotonically convergent and guaranteedly converge to a fixpoint, which is usually a very good one in practice. Our message passing algorithm transfers these desirable characteristics to a more general class of structured linear programs. Experiments on graph matching, MAP-MRF and multicut problems confirm the benefits of our apporach.

A particular problem in image segmentation is the segmentation of thin structures. Especially in biomedical images, the objects of interest often contain small scale elongated features that are very challenging to extract from noisy data. To solve these image segmentation tasks, connectivity constraints are a powerful extension to existing segmentation methods, and allow to preserve these thin structures. Application areas are, among others, blood vessel segmentation in angiography, quantification of cell protrusions in microscopy and also the three dimensional reconstruction of objects. I will present my recent work on connectivity constraints that are formulated on a tree that spans the whole image. The feasible set of this special instance of topological constraints form a convex cone and the resulting segmentation problem can be solved efficiently with a recent primal-dual algorithm for continuous convex optimization.