A paper with the title “Adapting the Segment Anything Model During Usage in Novel Situations” by Robin Schön, Julian Lorenz, Katja Ludwig and Rainer Lienhart has been accepted at the workshop for “Efficient Large Vision Models (eLVM)“. The workshop will be held jointly with the CVPR 2024 in Seattle. The paper presents a method for adapting the Segment Anything Model (SAM) during test time without the aid of additional training data. Instead, the method uses information with is generated during usage in order to generate pseudo labels.

The paper "A Review and Efficient Implementation of Scene Graph Generation Metrics" by Julian Lorenz, Robin Schön, Katja Ludwig, and Rainer Lienhart is accepted at the Workshop on Scene Graphs and Graph Representation Learning at CVPR 2024.

 

The authors review existing scene graph generation metrics and provide precise definitions that were lacking in this field. Additionally, they introduce an efficient and easy to use python package that implements all discussed metrics. To improve comparability of new scene graph generation methods, the authors provide a benchmarking service that enables an easy evaluation of scene graph generation models.

 

More information can be found here: https://lorjul.github.io/sgbench/

Lehrstuhl für Maschinelles Lernen und Maschinelles Sehen (Prof. Dr. Rainer Lienhart)

 

Bewerbungsschluss: 10. Mai 2024

We are always looking for excellent researchers who are passionate about research and a PhD. The research focus of our chair is machine learning and perception (vision/audio/other sensor modalities). A current research topic is, for instance, "Lifelong and continuous learning in single and multi-agent systems with sporadic human feedback“. Apply (f/m/d) with CV and previous study grades.

The paper titled "Towards Learning Monocular 3D Object Localization Using the Physical Laws of Motion" by Daniel Kienzle, Julian Lorenz, Katja Ludwig and Rainer Lienhart was accepted to the International Conference on 3D Vision (3DV) 2024. The paper describes a new method for localizing objects in 3D without the need for 3D ground truth. Instead, the method uses knowledge of physical laws to learn the task.

 

See https://kiedani.github.io/3DV2024/ for additional information on the paper.

The paper “Haystack: A Panoptic Scene Graph Dataset to Evaluate Rare Predicate Classes” by Julian Lorenz, Florian Barthel, Daniel Kienzle, and Rainer Lienhart is accepted at the First ICCV Workshop on Scene Graphs and Graph Representation Learning (SG2RL). The authors present Haystack, a new dataset for scene graph generation that tackles current shortcomings when evaluating with current scene graph datasets. Most notably, Haystack contains rare predicate classes and explicit negative annotations. Only through these properties can rare relationships be reliably evaluated. Based on the design of Haystack, the authors introduce three new scene graph metrics that can be used to gain more detailed insights about the prediction of rare predicate classes.

A paper with the title "Impact of Pseudo Depth on Open World Object Segmentation with Minimal User Guidance" by Robin Schön, Katja Ludwig and Rainer Lienhart has been accepted to the  2nd Workshop on Learning with Limited Labelled Data for Image and Video Understanding at the CVPR 2023. In this paper, the authors examine the effect of pseudo depth maps on the segmentation of object types which have not been present in the training data. The targeted objects are indicated by the means of coordinates on the object surfaces. In order to avoid a dependency ground truth depth maps, the depth maps are predicted by networks.

The paper with the title "All Keypoints You Need: Detecting Arbitrary Keypoints on the Body of Triple, High, and Long Jump Athletes" by Katja Ludwig, Julian Lorenz, Robin Schön and Rainer Lienhart is accepted at the 9th International Workshop on Computer Vision in Sports (CVsports) at CVPR 2023. In this paper, the authors detect arbitrary keypoints on the body of triple, high, and long jump athletes by extending previous methods with detections on the hands, feet, heads, elbows, and knees. Different representations regarding the head and the network input are evaluated in the paper.

The paper with the title "Detecting Arbitrary Keypoints on Limbs and Skis with Sparse Partly Correct Segmentation Masks" from Katja Ludwig, Daniel Kienzle, Julian Lorenz and Rainer Lienhart is accepted for the workshop Computer Vision for Winter Sports on the IEEE/CVF Winter Conference on Applications in Computer Vision (WACV) 2023. In this paper, the authors describe how to detect arbitrary keypoints on the limbs and skis of ski jumpers. Only a few, partly correct segmentation masks are necessary in the dataset for the presented method.

Das Paper mit dem Titel "Pseudo-Label Noise Suppression Techniques for Semi-Supervised Semantic Segmentation" von Sebastian Scherer, Robin Schön und Rainer Lienhart wurde für die British Machine Vision Conference (BMVC) 2023 akzeptiert. In diesem Paper beschreiben die Autoren eine Methode die es ermöglich, den Bedarf an großen gelabelten Datensätzen zu verringern, indem nicht gelabelte Daten in das Training einbezogen werden. Als Anwendung verwenden die Autoren die menschlicher Posenschätzung sowie die semantische Segmentierung, wobei besonderes letzteres interessant ist, da hier die Annotation von Daten äußerst zeitaufwendig ist.

Daniel Kienzle, Julian Lorenz, Katja Ludwig, Robin Schön and Rainer Lienhart from the chair for Machine Learning and Computer Vision achieved the first place in the STOIC challenge. The goal of the challenge was to predict the severe outcome of COVID-19 one month ahead using CT scans. To this end, the researchers employed convolutional neural networks and transfer learning on various tasks. The challenge was organized by Assistance Publique – Hôpitaux de Paris, Radboud University Medical Center, and Amazon Web Services.

 

The paper "Uplift and Upsample: Efficient 3D Human Pose Estimation with Uplifting Transformers" by Moritz Einfalt, Katja Ludwig and Rainer Lienhart is accepted at IEEE/CVF Winter Conference on Applications in Computer Vision (WACV) 2023. In this paper, the authors present a method to drastically reduce the computational complexity of 3D Human Pose Estimation with Transformers while maintaining smooth and precise 3D motion sequences.