Egocentric Room Location Classification Using Deep Neural Network Measuring Uncertainty

The rise of mobile applications in Virtual Reality (VR) and Augmented Reality (AR), particularly those using Head-Mounted Displays (HMDs), underscores the need to understand egocentric perspectives. This paper addresses the room-level localization challenge-identifying the room a user is in from an egocentric image-by framing it as a classification problem with a deep neural network. While deep learning has achieved remarkable success in conventional image classification, room classification from egocentric images introduces unique challenges due to variability and ambiguity in the user's perspective. Unlike typical datasets that provide clear visual data, egocentric views often lack sufficient detail, making uncertainty estimation crucial for achieving accurate results. Our approach not only advances egocentric localization but also holds potential for improving navigation and context-aware applications in AR/VR environments. We propose a novel strategy for uncertainty estimation and validate it with a custom dataset. Experimental results reveal significant performance improvements, achieving near-perfect accuracy by effectively managing ambiguous samples.