From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes

TL;DR

Anywhere3D-Bench introduces a holistic benchmark for multi-level 3D visual grounding, extending beyond object localization to area, space, and parts with 2,886 referring expression–3D bounding box pairs across 276 scenes. It evaluates LLMs, MLLMs, and 3D grounding models using GPT-4o-generated expressions refined by humans, revealing that space- and part-level grounding remain the most challenging even for top models (≈$30$–$40\%$ accuracy) and that visual-language models generally surpass 3D grounding specialists on space-level tasks. The results underscore a critical gap in current models’ ability to reason about 3D space from 2D inputs and motivate future work on structured reasoning, data augmentation, and embodied 3D instruction execution. The benchmark provides a rigorous, multi-granular testbed to push forward spatial understanding in 3D vision-language systems and informs practical applications in AR/VR and embodied AI.

Abstract

3D visual grounding has made notable progress in localizing objects within complex 3D scenes. However, grounding referring expressions beyond objects in 3D scenes remains unexplored. In this paper, we introduce Anywhere3D-Bench, a holistic 3D visual grounding benchmark consisting of 2,886 referring expression-3D bounding box pairs spanning four different grounding levels: human-activity areas, unoccupied space beyond objects, individual objects in the scene, and fine-grained object parts. We assess a range of state-of-the-art 3D visual grounding methods alongside large language models (LLMs) and multimodal LLMs (MLLMs) on Anywhere3D-Bench. Experimental results reveal that space-level and part-level visual grounding pose the greatest challenges: space-level tasks require a more comprehensive spatial reasoning ability, for example, modeling distances and spatial relations within 3D space, while part-level tasks demand fine-grained perception of object composition. Even the best-performing models, Google Gemini-2.5-Pro and OpenAI o3, achieve just around 30% accuracy on space-level tasks and around 40% on part-level tasks, significantly lower than its performance on area-level and object-level tasks. These findings underscore a critical gap in current models' capacity to understand and reason about 3D scenes beyond object-level semantics.

Figures (32)

• Figure 1: Multi-level Visual grounding in 3D Scenes: area, space, object, and part. Examples illustrate visual grounding of daily life expressions, from functional areas for collaborative study, to placing a cup on a nightstand in unoccupied space, referring to an object via its spatial distance from the armchair, or moving part of an object such as pulling out a drawer from a cabinet.
• Figure 2: Multi-level visual grounding (Part, Space, Object, Area) with distinct expression types. Emphasized segments highlight phrases aligned with their respective expression type.
• Figure 3: Data generation pipeline of Anywhere3D-Bench. We design specific prompts for different types of expression in four grounding levels. Human annotators are required to annotate the 3D bounding box and refine the GPT-generated expression until they precisely match.
• Figure 4: Results on different types of expressions on Space Level.
• Figure 5: Results on different types of expressions on Object Level.