Location-Reasoning for 3D Question Answering in Large Language Models

Large Language Models significantly improve 3D scene understanding, especially in 3D Question Answering. However, reasoning about positional relationships between objects remains a challenge. Existing methods often overlook the full impact of positional relations or rely on artificial prior knowledge. We propose a location-reasoning approach for 3D Question Answering, allowing LLM to infer both local and global positional relationships within scenes. Our experiments show that incorporating positional relation inference enhances the performance of 3D Question Answering tasks.