ScanReason: Empowering 3D Visual Grounding with Reasoning Capabilities

Chenming Zhu^1,2, Tai Wang², Wenwei Zhang², Kai Chen², Xihui Liu^1*

¹The University of Hong Kong, ²Shanghai AI Laboratory

We propose a new task: 3D Reasoning Grounding task, focusing the reasoning grounding in 3D scene. Given a 3D scene and an implicit free-form query, 3D reasoning grounding requires the model to predict the 3D bounding boxes of the target objects as well as the textual answers and explanations.

ScanReason is the first comprehensive and hierarchical 3D reasoning grounding benchmark. We define 5 types of questions depending on which type of reasoning is required: Spatial reasoning and function reasoning require fundamental understanding of the 3D physical world, focusing on objects themselves and inter-object spatial relationships in a 3D scene respectively, and logistic reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications.

Abstract

Although great progress has been made in 3D visual grounding, current models still rely on explicit textual descriptions for grounding and lack the ability to reason human intentions from implicit instructions. We propose a new task called 3D Reasoning Grounding and introduce a new benchmark ScanReason which provides over 10K question-answer-location pairs from five reasoning types that require the synerization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. Extensive experiments on the proposed benchmark validate the effectiveness of our proposed approach.

Method

We propose ReGround3D consisting of a visual-centric reasoning module and a 3D grounding module with geometry-enhanced look-back. The visual-centric reasoning module performs joint reasoning of language instruction and visual scene, and predicts a special <LOC> token representing the grounding information. The 3D grounding module looks back to the original 3D scene with comprehensive geometry information and fine-grained details. It takes the hidden embedding of the <LOC> token containing grounding-related information from the 3D features, and eventually predicts the 3D locations of the target objects.

Furthermore, we propose Chain-of-Grounding mechanism (CoG), a chain of interleaved reasoning and grounding steps, to further synergize the grounding and reasoning capability for the 3D reasoning gruonding task.

Benchmark Visualization