Agent3D-Zero: An Agent for Zero-shot 3D Understanding

TL;DR

<3-5 sentence high-level summary> Agent3D-Zero presents a zero‑shot framework for 3D scene understanding by leveraging vision‑language models to reason over multiple 2D views. It introduces Set‑of‑Line Prompting to convert BEV observations into controllable prompts that guide viewpoint planning and pose estimation, enabling active multi‑view analysis without 3D data training. The method demonstrates strong zero‑shot performance on 3D QA and enables 3D captioning, dialogue, and perception tasks by back‑projecting 2D segmentations into 3D space, with notable ablations validating the benefit of more viewpoints and denser prompts. Collectively, the work reduces reliance on annotated 3D data and finetuning, illustrating a scalable path for 3D understanding using foundation models.

Abstract

The ability to understand and reason the 3D real world is a crucial milestone towards artificial general intelligence. The current common practice is to finetune Large Language Models (LLMs) with 3D data and texts to enable 3D understanding. Despite their effectiveness, these approaches are inherently limited by the scale and diversity of the available 3D data. Alternatively, in this work, we introduce Agent3D-Zero, an innovative 3D-aware agent framework addressing the 3D scene understanding in a zero-shot manner. The essence of our approach centers on reconceptualizing the challenge of 3D scene perception as a process of understanding and synthesizing insights from multiple images, inspired by how our human beings attempt to understand 3D scenes. By consolidating this idea, we propose a novel way to make use of a Large Visual Language Model (VLM) via actively selecting and analyzing a series of viewpoints for 3D understanding. Specifically, given an input 3D scene, Agent3D-Zero first processes a bird's-eye view image with custom-designed visual prompts, then iteratively chooses the next viewpoints to observe and summarize the underlying knowledge. A distinctive advantage of Agent3D-Zero is the introduction of novel visual prompts, which significantly unleash the VLMs' ability to identify the most informative viewpoints and thus facilitate observing 3D scenes. Extensive experiments demonstrate the effectiveness of the proposed framework in understanding diverse and previously unseen 3D environments.

Figures (4)

• Figure 1: Illustration of (a) finetuning-based paradigm and (b) our proposed zero-shot paradigm. The finetuning-based paradigm exploits an external 3D perceiver, and finetunes it with a frozen LLM. On the contrary, our proposed zero-shot paradigm is simple and efficient, directly utilizing the VLM to actively select and interpret multiple observing views for zero-shot 3D task resolution.
• Figure 2: System Overview of Agent3D-Zero. The upper segment illustrates our viewpoints-selection progress. We initiate the process by overlaying grid lines and tick marks on the Bird's Eye View (BEV) images, constituting the prompt along with a scene type description. This prompt guides the Vision Language Model (VLM) to retrieve camera poses for images observing the 3D scene. The lower section demonstrates the versatility of Agent3D-Zero, showcasing its proficiency in addressing various 3D reasoning and perception tasks through strategic prompting and tool utilization.
• Figure 3: Visualization of 3D Scene Caption and Task Decomposition of Agent3D-Zero. The top part presents the raw 3D scan and some of the images selected from different viewpoints. We show examples of 3D Scene Caption and Task Decomposition at the bottom.
• Figure 4: Visualization of navigation in real world. The top section introduces the navigation task and provides an overview of an office setting. Subsequent rows feature observations of the environment, with GPT-4v-generated instructions on the right. The visualization concludes with the VLM successfully locating the printer, thereby accomplishing the task in an unfamiliar environment.