informatik-vortraege Mai 2024

informatik-vortraege@lists.rwth-aachen.de

3 Teilnehmer
4 Diskussionen

Einladung: Informatik-Oberseminar Sabarinath Mahadevan

von Mahadevan, Sabarinath

+********************************************************************** * * * Einladung * * * * Informatik-Oberseminar * * * +********************************************************************** Zeit: Freitag, 24. Mai 2024, 10.30 Uhr Ort: Raum 025, Mies-van-der-Rohe Str. 15 (UMIC Gebäude) Der Vortrag findet hybrid statt. Zoom: https://rwth.zoom-x.de/j/64536789621?pwd=d2I5Y1VDM2xscVVDbTF0ZnJVWUUvZz09 Meeting ID: 645 3678 9621, Passcode: 998171 Referent: Sabarinath Mahadevan, M.Sc. Lehrstuhl Informatik 13 Thema: The Many Facets of Object Segmentation in Images and Videos Abstract: Segmentation is an important task in computer vision where the pixels belonging to a region of interest, that often share similar characteristics, have to be separated out, and assigned a unique label. This task is relevant in both the image and video domains, and finds direct applications in a wide range of fields including but not limited to autonomous driving, robotics, image editing, and surveillance. For both images and videos, the task of segmentation has various flavours that are often highly related and equally challenging. With the advent of deep learning, modern computer vision algorithms are able to leverage large amounts of available data to achieve impressive performance for many of these segmentation problems. However, such algorithms are often catered either towards segmenting a pre-defined set of object categories, or to some specific sub-domains, and require to be adapted for out-of-domain tasks by training these methods on additional domain-specific data, that is expensive to annotate. As a result, we need algorithms that generalise well to data from unseen domains and also to new task settings in addition to having methods that can efficiently annotate data. The first part of this thesis focusses on efficiently annotating objects using Interactive Segmentation, where the goal is to segment and refine objects in an image using user clicks. Here, I'll present two of our research works that advances the state-of-the-art in the interactive segmentation domain. The first among these, ITIS, develops a novel iterative training strategy in which clicks are added iteratively during training based on the error regions in the network predictions. The iterative training strategy aligns the simulated user-click patterns that are used during training with the actual click patterns that the network would encounter at test time. While this strategy is effective in reducing the number of clicks required to annotate objects, ITIS can only segment one object at a time due to the limitations of its network architecture. We address this problem in our subsequent work called DynaMITe, where we formulate user clicks as spatiotemporal sequences, and develop a novel Transformer based formulation that can process such a sequence and encode them into relevant object or region descriptors. These descriptors are then used to generate the relevant instance segmentation masks. Unlike previous methods, our architecture can process clicks for multiple objects at once, and correspondingly predicts non overlapping segmentation masks without any post-processing. In the second part of this talk, I'll focus on end-to-end video segmentation networks based on 3D convolutions, and present our work STEm-Seg. STEm-Seg is a a bottom-up end-to-end approach for instance segmentation on videos, which uses a partially 3D network to learn spatio-temporal embeddings that can be clustered into instance tubes based on the predicted clustering parameters. Our method is very generic and can be applied to a wide range of instance segmentation tasks in videos. Finally, I’ll present our latest work, Point-VOS, where we show that video segmentation models can learn from spatiotemporally sparse point annotations instead of dense per-object mask annotations. We also present an efficient point-wise annotations scheme, and use it to annotate two large-scale video datasets with associated language expressions. We also present a new Point-VOS benchmark and the corresponding baselines, and show that on our point annotations can be used to achieve results close to state-of-the-art models that use dense mask supervision. Additionally, we evaluate models that connect vision and language on the VNG task, and correspondingly show that our data helps in improving their performance.

1 Monat, 1 Woche

UnRAVeL Survey Lecture - Today!

von Kassing, Jan-Christoph

Dear all, this is a reminder for Martin Grohe's talk with the title "The Complexity of Constraint Satisfaction" taking place today (16.05) at 12:30 in the B-IT room 5053.2. Please find the details below --- Abstract --- The objective of a constraint satisfaction problem (CSP) is to assign values to variables subject to constraints on the values. Obviously, this is a very general type of problem, and it is not surprising that many algorithmic problems in various application domains can be described as CSPs. It is neither surprising that, in general, CSPs are computationally hard. Considerable efforts have been made to precisely understand the complexity of CSPs, with the goal of identifying tractable restrictions and, ultimately, determining the boundary between tractable and intractable CSPs. A deep theory based on methods from logic, combinatorics, and universal algebra has evolved around these questions. In this survey talk, I will give an overview of the main results and techniques of this theory, and I will highlight some current research topics and open questions. ---------------- Part of the programme of the research training group UnRAVeL is a series of lectures on the topics of UnRAVeL’s research thrusts algorithms and complexity, verification, logic and languages, and their application scenarios. Each lecture is given by one of the researchers involved in UnRAVeL. This years topic is "UnRAVeL - New Ideas!". In these lectures, UnRAVeL professors will discuss current research as well as highlight open problems and offer a perspective on potential future directions. All interested doctoral researchers and master students are invited to attend the UnRAVeL lecture series 2024 and engage in discussions with researchers and doctoral students. We are looking forward to seeing you at the lectures. Kind regards, Jan-Christoph for the organisation committee

1 Monat, 2 Wochen

Einladung: Informatik-Oberseminar Barna Zajzon

von Abigail Morrison

+********************************************************************** * * * Einladung * * * * Informatik-Oberseminar * * * +********************************************************************** Zeit: 08.05.2024 um 09:00 Ort: Informatikzentrum, Raum 9222 in E3 Referent: Barna Zajzon, M.Sc. Lehrstuhl Informatik 3 Thema: Sequential information processing in modular spiking networks Abstract: Established paradigms formulate many cognitive processes in terms that involve the manipulation of sequentially organized time-discrete (symbolic) representations. This underscores two basic functional requirements that cortical circuits must fulfill: the ability to create suitable representations from a highly volatile and noisy environment; and the capacity to process, and learn from, their spatio-temporal structure. Combining software tools, simulation studies and theoretical analysis, this thesis entails a series of research projects with the shared goal of disentangling how modular structures enable neural circuits to learn and process sequential information in an efficient and reliable manner. The first part analyses the characteristics of state representations in modular spiking networks and the architectural and dynamical constraints that influence the system's ability to retain, transfer and integrate stimulus information in the presence of noise. It explores the novel hypothesis that modular topographic maps, a pervasive anatomical feature of the cortex, may provide a structural scaffold for sequential denoising of stimulus representations. The second part of this work is dedicated to investigating existing, biologically detailed models of sequence processing. If we are to harvest the knowledge within these models and arrive at a deeper mechanistic understanding of the involved phenomena, it is critical that the models and their findings are accessible, reproducible, and quantitatively comparable. We illustrate these aspects through a replication study and then lay the initial steps towards a conceptual and practical, theoretically-grounded framework for benchmarking and comparison of biophysical sequence learning models. Es laden ein: die Dozentinnen und Dozenten der Informatik -- Prof. Dr. Abigail Morrison Institute for Advanced SImulation (IAS-6) Jülich Research Center & Computer Science 3 - Software Engineering RWTH Aachen http://www.fz-juelich.de/inm/inm-6 http://www.se-rwth.de Office: +49 2428 8097504 Fax # : +49 2461 61-9460 Pronouns: she/her ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------ Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens ------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------

1 Monat, 3 Wochen

UnRAVeL Survey Lecture - Today!

von Kassing, Jan-Christoph

Dear all, this is a reminder for Jürgen Giesl's talk with the title "Termination and Complexity Analysis of (Probabilistic) Programs: Results and Future Work" taking place today at 12:30 in the B-IT room 5053.2. Please find the details below --- Abstract --- We give an overview on our work on developing techniques for automated termination and complexity analysis for different kinds of programs. While automated tools for the analysis of non-probabilistic programs are already very powerful, we will indicate several research problems that have to be solved in order to obtain techniques of similar power for probabilistic programs. ---------------- Part of the programme of the research training group UnRAVeL is a series of lectures on the topics of UnRAVeL’s research thrusts algorithms and complexity, verification, logic and languages, and their application scenarios. Each lecture is given by one of the researchers involved in UnRAVeL. This years topic is "UnRAVeL - New Ideas!". In these lectures, UnRAVeL professors will discuss current research as well as highlight open problems and offer a perspective on potential future directions. All interested doctoral researchers and master students are invited to attend the UnRAVeL lecture series 2024 and engage in discussions with researchers and doctoral students. We are looking forward to seeing you at the lectures. Kind regards, Jan-Christoph for the organisation committee

2 Monate

2024

2023

2022

2021

2020

2019

2018

informatik-vortraege Mai 2024