Enhancing Scene Identification Performance through Hierarchical Classification, Context-Aware Analysis, and Active Learning Operations

Abstract

The increasing demand for intelligent systems in various applications, such as smart homes, autonomous vehicles, and surveillance, has underscored the need for robust and efficient Scene Identification techniques. Accurate Scene Identification plays a critical role in understanding complex environments, enabling more effective decision-making and interaction with the environment. However, existing methods often suffer from high computational costs, low adaptability to new scenarios, and limited ability to capture context and object relationships. In this paper, we address these limitations by lodging a novel Scene Identification model that combines Hierarchical Scene Identification, Context-Aware Detection, and Active Learning techniques. Our approach capitalizes on the inherent hierarchical nature of objects, leveraging a two-stage object detector to categorize objects first into broad categories and then into specific objects. We introduce an efficient Graph Neural Network (GNN) to capture contextual information between objects, enhancing the detection process's accuracy and robustness. Active Learning is applied to actively query labels for uncertain instances, significantly reducing the manual labeling effort and improving model’s performance levels. The proposed model demonstrates remarkable improvements in various evaluation metrics compared to existing methods.