Toward High-Level Visual Content Interpretation and Annotation for Sport Events

This paper presents a novel framework for visual content interpretation and aims to generate meaningful descriptions for visual content based on the aggregate information of the detected primitive objects, spatial relations, and specific relevant prior knowledge to aid visual content interpretation. The main contributions of this paper include: (1) a novel approach to generate semantic descriptions for visual data at natural language level whereas the state of the art frameworks perform just simple object labeling which is not informative; (2) an integration of the detected primitive objects, spatial relations and contextual knowledge to detect the action scene (competition phase) in sport events e.g. flying action in a pole vault event. The experimental results show that the presented approach can discover semantically meaningful visual content descriptions and recognize sport event and action in the visual data efficiently.