Physically Interpretable Probabilistic Domain Characterization

• DOI: 10.48550/arxiv.2411.14827
• Publication Date: 2024-11-22
• Authors: Anaïs Halin; Sébastien Pierard; Renaud Vandeghen; Benoît Gérin; Maxime Zanella; Martin Colot; Jan Held; Anthony Cioppa; Emmanuel Jean; Gianluca Bontempi; Saïd Mahmoudi; Benoı̂t Macq; Marc Van Droogenb...
• Keywords: Characterization

Characterizing domains is essential for models analyzing dynamic environments, as it allows them to adapt evolving conditions or hand the task over backup systems when facing outside their operational domain. Existing solutions typically characterize a domain by solving regression classification problem, which limits applicability they only provide limited summarized description of In this paper, we present novel approach characterization characterizing probability distributions. Particularly, develop method predict likelihood different weather from images captured vehicle-mounted cameras estimating distributions physical parameters using normalizing flows. To validate our proposed approach, conduct experiments within context autonomous vehicles, focusing on predicting distribution This characterized (absolute characterization) and arbitrarily predefined (relative characterization). Finally, evaluate whether system can safely operate in target comparing multiple source where safety has already been established. holds significant potential, accurate prediction effective adaptation are crucial adjust environmental conditions.