Open World: Generalize and Recognize Novelty
Tatiana Tommasi - Politecnico di Torino
Navigating the world, autonomous agents will inevitably encounter objects belonging to unfamiliar semantic classes. Naïvely assigning labels from a closed set of categories may lead to erroneous decisions, posing safety risks. In this scenario, it is crucial to design models that acknowledge their ignorance when faced with new stimuli. This ability is the focal point of out-of-distribution (OOD) detection, specifically semantic novelty detection. While numerous works explore OOD detection algorithms, many are tested on small datasets, neglecting settings crucial for real-world applications. This talk will shed light on these overlooked aspects and explore strategies to leverage foundational models for semantic novelty detection in both 2D (images) and 3D (point-clouds) data.
Tatiana Tomasi is Associate Professor at the Department of Control and Computer Engineering of Politecnico di Torino and Affiliated Researcher at the Italian Institute of Technology. Her research area includes computer vision and machine learning. The main topic of her work is knowledge transfer, domain adaptation and object categorization using multimodal information. I'm also interested in natural language processing, medical image analysis and robotics.
Monday December 11th, 16:00
Room 705, DIMA, Via Dodecaneso 35