Contextual Object Classification via Geo-Semantic Scene Graphs

This paper addresses the limitations of traditional object recognition systems by emphasizing the importance of contextual information. It introduces a novel framework using Geo-Semantic Contextual Graphs (GSCG) to represent scenes and a graph-based classifier to leverage this context. The results demonstrate significant improvements in object classification accuracy compared to context-agnostic models, fine-tuned ResNet models, and even a state-of-the-art multimodal LLM. The interpretability of the GSCG approach is also a key advantage.

• •Contextual information is crucial for object recognition.
• •The Geo-Semantic Contextual Graph (GSCG) provides a rich, structured representation of scenes.
• •A graph-based classifier effectively leverages contextual information.
• •The proposed approach significantly outperforms existing methods, including LLMs, in object classification accuracy.
• •The GSCG approach offers inherent interpretability.