This benchmark reframes evaluation of 3D visual grounding around the linguistic diversity of referring prompts rather than raw dataset scale. The authors define a framework that categorizes grounding prompts by linguistic phenomena (such as negation, coarse-grained references, anaphoric reference resolution, and varied syntax) and assemble a compact diagnostic dataset that deliberately spans these patterns. Students learn how a targeted, linguistically-analyzed evaluation set exposes failure modes of open-vocabulary grounding models on out-of-distribution language that large homogeneous benchmarks miss, and how to reason about coverage of natural-language variation when designing vision-language benchmarks.
This computer-vision and computational-linguistics paper addresses 3D visual grounding, the task of localizing the entities in a 3D scene that are referred to by a natural-language description. Obser…