The Referlt3D dataset is suitable for 3D visual grounding and 3D dense captioning tasks, including 2 large-scale visio-linguistic complementary datasets:

1. Sr3D dataset: contains 83.5K template-based sentences, which uses spatial relations between fine-grained object classes to locate the objects referenced in the scene;
2. Nr3D dataset: includes 41.5K naturally occurring, free-form sentences collected by deploying a 2-player game in a 3D scene.