April 23, 2025 | 5:30 – 8:30 PM
Date and Time
April 23, 2025 from 5:30 PM to 8:30 PM
Location
Impact Hub Stuttgart, Quellenstraße 7a Stuttgart
Porsche AG
When using Advanced Driver Assistance Systems (ADAS), drivers often take over control of the active driving function to adjust its behavior to their own personal preferences. These takeovers can be used as feedback for the optimization and individualization of the driving function, if interpreted correctly.
In this presentation, a bottom-up data-driven approach is highlighted how ADAS optimization potential can be derived from real driving data. The underlying methodologies are developed by using the data of a test group study featuring naturalistic driving data. Then, a prototypical self-learning driving function is proposed, and its performance is evaluated in a second test group study.
FIZ Karlsruhe
One major challenge to date in the field of Document Processing is transforming analogue documents into computer-readable formats. Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR) techniques are traditional methods for this transformation. Despite progress in text recognition through OCR and HTR, this issue remains largely unresolved, particularly regarding historical documents stored in archives, due to visual complexities such as overlapping areas, paper degradation, and ink fading. In the context of the project “Wiedergutmachung”, we propose a pipeline to address the issue of text type heterogeneity in single document images by decomposing the document into its constituent text types—handwritten and machine-printed text, to enhance text recognition accuracy by utilising appropriate models for each text layer, in order to improve the quality of final transcripts.
University of Freiburg
Training high-quality CLIP models typically requires enormous datasets, which limits the development of domain-specific models and drives up training costs. In this work, we show that by employing image search based on knowledge graphs, a robust CLIP model can be trained from scratch with considerably less data. Specifically, we introduce the EntityNet dataset comprising 33M images paired with 46M text descriptions. We train an expert foundation model using a subset of 10M images of living organisms, and we train a generic CLIP model on the full dataset.
Voxel51
This talk covers methods to label images in the main computer vision tasks:
We look at combining zero-shot classifiers, like CLIP, with active learning. We will discuss key implementation details such as:
Join the AI and ML enthusiasts who have already become members
The goal of the AI, Machine Learning, and Computer Vision Meetup network is to bring together a community of data scientists, machine learning engineers, and open source enthusiasts who want to share and expand their knowledge of AI and complementary technologies. If that’s you, we invite you to join the Meetup closest to your timezone.