ZESS Lecture Series: Talks in Visual Computing

In the coming weeks, the ZESS Lecture Series will host three guest talks from the domain of Visual Computing by researchers from MPI Saarbrücken, TU Darmstadt and ETH Zurich.

 

Anna Kukleva (MPI Saarbrücken): Advancing Image and Video Recognition with Less Supervision

14 March 202, 14-15 h in H-C 7327

Deep learning has become an essential component of modern life, transforming various tasks across multiple domains such as entertainment, education, and autonomous driving. However, the increasing demand for data to train models for emerging tasks poses significant challenges. Deep learning models heavily rely on high-quality labeled datasets, yet obtaining comprehensive supervision is resource-intensive and can introduce biases. Therefore, this talk explores strategies to mitigate the need for full supervision and reduce data acquisition costs. The first part of the discussion focuses on self-supervised and unsupervised learning methods, which enable learning without explicit labels by leveraging inherent data structures and injecting prior knowledge for robust data representations. The second part of the presentation discusses strategies such as minimizing precise annotations in multimodal learning, allowing for effective utilization of correlated information across different modalities. Moreover, we discuss open-world scenarios, proposing novel setup and method to adapt vision-language models to the new domains. Overall, this research contributes to understanding learning dynamics and biases present in data, advancing training methods that require less supervision.

 

Dr Simone Schaub-Meyer (TU Darmstadt): Efficient and Understandable Neural Networks for Image and Video Analysis

21 March 2024, 14-15 h in H-C 7327

Recent developments in deep learning have led to significant advances in many areas of computer vision. However, the success of these methods often depends on having a well-defined task corresponding training data, and measuring success by improved task specific accuracy. However, in order to apply new methods in the real world, other aspects become relevant as well, such as required labelled data, computational requirements, as well as, especially in safety critical scenarios, how trust worthy a model is.

In my talk, I will first discuss how motion in videos can be used to learn representations in an unsupervised way as well as methods to efficiently handle higher resolution data. In the second part, I will show how attribution maps, which help to gain a better understanding of the predictions, can be obtained efficiently, as well as how we can evaluate them.

 

Dr Alexandra Diehl (ETH Zurich): Visualization Research for Characterization of High Impact Weather Events (HIWE)

25 March 2024, 14-15 h in H-C 7327

The talk will be held at a later date.

The efficient analysis, decision making, and communication of high-impact weather events (HIWE) and their associated risks is a challenging and ongoing research topic for meteorologists and computer scientists. The challenge is due, in part, to the inherent unpredictability of weather and the difficulty of quantifying its risk and communicating its uncertainties.

In this talk, I will summarize my recent contributions to the research on efficient visualization tools for the analysis and communication of weather forecasts and the characterization of HIWEs. I will also present my current efforts in citizen data analysis and discuss the open challenges for efficient communication of severe weather events through citizen participatory science.

 

Dr Jing Ren (ETH Zurich): Shape matching and map space exploration via functional maps

26 March 2024, 13-14 h in H-C 7327

Computing semantically meaningful correspondences between shapes is an important yet challenging task in computer graphics and geometry processing, with a wide range of applications including deformation, motion, and texture transfer , as well as 3D shape retrieval, shape analysis and exploration, among others. The functional map framework emerges as a powerful solution to this shape matching problem. Rather than directly attempting to establish point-wise correspondences, it operates by matching functions defined on shapes.

This presentation delves into techniques build upon functional map framework. Specifically, we will explore (1) how to design various regularizers to resolve orientation ambiguity or promote matching accuracy; (2) how to refine a given point-wise map effectively and efficiently; (3) and how to formulate and structure the map space to explore all possible high-quality maps between a shape pair.