VLM-Grounder: A VLM Agent for Zero-Shot 3D Visual Grounding

TL;DR

The paper tackles zero-shot 3D visual grounding by introducing VLM-Grounder, a GPT-4V-based agent that grounds targets directly from 2D image sequences without relying on 3D point clouds or priors. It couples dynamic image stitching, a grounding-and-feedback loop, and a multi-view ensemble projection to estimate 3D bounding boxes from 2D masks, supported by a novel Visual-Retrieval Benchmark to study stitching effects. Empirical results on ScanRefer and Nr3D show state-of-the-art zero-shot performance and competitive standing against some supervised methods, while clearly exposing limitations due to sensor noise and 3D projection inaccuracies. The work highlights the viability of VLM-based agents for 3D grounding and provides a framework for improving efficiency through local deployment and better 2D detectors.

Abstract

3D visual grounding is crucial for robots, requiring integration of natural language and 3D scene understanding. Traditional methods depending on supervised learning with 3D point clouds are limited by scarce datasets. Recently zero-shot methods leveraging LLMs have been proposed to address the data issue. While effective, these methods only use object-centric information, limiting their ability to handle complex queries. In this work, we present VLM-Grounder, a novel framework using vision-language models (VLMs) for zero-shot 3D visual grounding based solely on 2D images. VLM-Grounder dynamically stitches image sequences, employs a grounding and feedback scheme to find the target object, and uses a multi-view ensemble projection to accurately estimate 3D bounding boxes. Experiments on ScanRefer and Nr3D datasets show VLM-Grounder outperforms previous zero-shot methods, achieving 51.6% Acc@0.25 on ScanRefer and 48.0% Acc on Nr3D, without relying on 3D geometry or object priors. Codes are available at https://github.com/OpenRobotLab/VLM-Grounder .

Figures (11)

• Figure 1: Comparison between LLM-based methods and VLM-Grounder.
• Figure 2: An overview of VLM-Grounder. VLM-Grounder analyzes the user query and dynamically stitches image sequences for efficient VLM processing to locate the target image and object. A 2D open-vocabulary detection model and the Segment Anything Model generate a fine-grained mask, which is then projected using a multi-view ensemble strategy to obtain the 3D bounding box.
• Figure 3: Visual-Retrieval benchmark.
• Figure 4: Benchmark accuracy and request time for different stitching layouts and image counts.
• Figure 5: Failure cases of the VLM grounding module.