What are you looking for?

Direct Visual SLAM for Autonomous Systems - The reconstruction of the 3D world from moving cameras has seen enormous progress over the last couple of years. Already in the 2000s, researchers have pioneered algorithms which can reconstruct camera motion and sparse feature-points in real-time. In my talk, I will introduce direct methods for camera tracking and 3D reconstruction which do not require feature point estimation, which exploits all available input data and which recover dense or semi-dense geometry rather than sparse point clouds. Experimental results confirm that the direct approaches lead to a drastic boost in precision and robustness. I will present recent developments on Simultaneous Localization and Mapping (SLAM) using monocular and stereo cameras, inertial sensors and deep neural networks with applications to autonomous systems.

10 years of lies, damned lies, and TED Talks - Have you ever wondered what makes certain TED Talks go viral? Why you enjoy some of them and others much less? 10 years after his original analysis, Dr. Sebastian Wernicke (re-)analyzed all TED Talks over the last 15 years and conducted statistical analyses to identify what makes the most memorable ones. In a tongue-in-cheek analysis, he looks at delivery style, profession, talk duration, and many more interesting factors. See what makes the most effective TED Talk (according to the data)!

Extracting business value from half a billion daily user events - Avira is receiving half a billion user events per day, ranging from product usage events to campaign reactions to purchase and refund events. This stream of data is the foundation for various in-house data products like product health dashboards showing retention and stickiness of our products, and an experimentation framework for rapid AB testing. The fundamental question behind such behavioral data products is always "How is the customer using our product?". Generated insights enable our product owners and campaign managers to take actions that ultimately drive product and business growth. In this talk, we present a framework developed and used at Avira for behavioral analysis. The framework establishes an event data methodology providing methods to mine and explore event data. We will show different use cases, e.g., attribution models based on Markov models helping to find out relevant touch points leading to a con

Reproducible Data Science - Data science projects are generally complex, involve data cleaning, visualization, and machine learning pipelines, and should, in the end, be reproducible and presentable. In this talk, I will show some issues with making data science projects reproducible and discuss some solutions such as automation, good naming, and structure of folders and files, literate programming, using stable computing environments, etc.

Putting your AI models in production using the PyTorch - A favourite with researchers since it became public, over the last year PyTorch has gained excellent support for production deployment as well. We will take a tour of how to make the most of PyTorch and its integrated Just In Time compiler to train models fast and deploy them in various settings, from network serving to integration in larger non-Python applications, to mobile devices. Our main theme will be the practical aspects but we also take a peek under the hood to understand how PyTorch delivers these.

Driving fully autonomously in urban environment – which is the mission from AID – brings new challenges for the automotive industry. Cumbersome feature extraction and algorithm design which require years of experience in the field are increasingly often outperformed by Deep Learning models, so AI plays an important role in this field. This session is intended to give an overview of architectural approaches involved with robotaxis, especially on the sw stack enhanced with AI models, car compute platform and sw/hw integration.

More than eighty percent of an organization’s data has a location component and chances are it is being ignored thirty percent of the time. How can we turn this untapped data into value? Location Intelligence is the practice of collecting, enriching, managing and analyzing location data to inform optimized actions, decisions, and customer experiences. In this talk, I will overview the evolution of location data and discuss how to avoid common analytical pitfalls with some practical examples.

Safe Sex - The idea of synthetic data is to mimic the statistical properties of a real dataset without exposing individual entities but exchanging them for synthetic ones. All across the industry (e.g. banks, insurances, ...) this is in high demand due to increased privacy concerns (GDPR). Different approaches exist to generate synthetic data like PCA, autoencoders and generative models. The big challenge is to correctly represent the statistical properties while avoiding simply generating duplicates. Only then it is safe to use this procedure as the latter would expose protected information. In this presentation, we compare the performance of two generative models, namely variational autoencoders and GANs by evaluating how succesful the output is to train supervised and unsupervised models with respect to the real data. Furthermore, we employ the KNN-algorithm to examine the similarity between synthetic and real data and thus determine which algorithm generates entities to be exposed

The Computational Science of Real-time Influenza Forecasting - Seasonal influenza outbreaks cause substantial annual morbidity and mortality worldwide. Accurate forecasts of key features of influenza epidemics, such as the timing and severity of the peak incidence in a given season, can inform public health response to outbreaks. As part of ongoing efforts to incorporate real-time data and advanced computational methods into public health decision-making, the United States Centers for Disease Control and Prevention (CDC) has organized real-time seasonal influenza forecasting challenges since 2013. This talk will summarize the recent collaborative efforts by the FluSight Network consortium to improve real-time forecast accuracy through multi-model ensemble forecasting. We will discuss the process of building and evaluating ensemble approaches, each of which combines over 20 separate component model predictions into a single probabilistic forecast. During the Northern hemisphere winter i