BMVC Workshop
The workshop featured keynote presentations from distinguished speakers:
- Jiří Matas, Head of the Visual Recognition Group at the Center for Machine Perception, CTU Prague, presented "Robust Recognition with Image Decomposition", exploring the disentaglement and fusion of background and foreground information in image classification.
- Petra Bevandić, a Postdoctoral researcher at the University of Bielefeld, discussed "Addressing Incompatible Taxonomies for Effective Multi-Dataset Training" highlighting strategies for aligning taxonomies across diverse labeling policies of datasets.
- Toby Breckon, Professor at Durham University, delivered insights on "Object-wise Out of Distribution & Anomaly Detection Meets Real-World Complexity" applied to the hands on task of airport security controls.
- Robert Geirhos, Research Scientist at Google DeepMind, concluded with "OOD Generalization with Generative Models", demonstrating human-aligning characteristics of generative models as classifiers.
There was an array of contributed talks, covering topics such as unsupervised feature orthogonalization, hybrid video anomaly detection, and our collaborative work titled "Uncertainty and Prediction Quality Estimation for Semantic Segmentation via Graph Neural Networks" by Edgar Heinert, Stephan Tilgner, Timo Palm, and Matthias Rottmann. A highlight was the challenge session on Open-World Object Detection and Tracking, encouraging participants to develop AI models capable of detecting and tracking objects in diverse and unpredictable environments.
Besides that, we were happy to also present our work "AttEntropy: On the Generalization Ability of Supervised Semantic Segmentation Transformers to New Objects in New Domains" by Krzysztof Lis, Matthias Rottmann, Annika Mütze, Sina Honari, Pascal Fua, Mathieu Salzmann. This is a joint work of University of Wuppertal, EPFL Lausanne and Samsung AI Center Toronto.
