Scaling Diverse Language Generation for 3D Visual Grounding

The article introduces ViGiL3D++, a scalable method for 3D visual grounding (3DVG) that enhances the generation of diverse grounding queries by integrating constraint sampling from scene graphs with language generation from large language models (LLMs). This approach demonstrates improved diversity compared to existing datasets and enhances performance on multiple 3DVG benchmarks, while also highlighting the limitations of vision-language models (VLMs). This development is significant for practitioners as it addresses the challenge of generalizing spatial language understanding in 3D environments, which is crucial for the deployment of AI agents in real-world applications.