Toward General Object-level Mapping from Sparse Views with 3D Diffusion Priors

TL;DR

This work proposes a General Object-level Mapping system, GOM, which leverages a 3D diffusion model as shape prior with multi-category support and outputs Neural Radiance Fields (NeRFs) for both texture and geometry for all objects in a scene.

Abstract

Object-level mapping builds a 3D map of objects in a scene with detailed shapes and poses from multi-view sensor observations. Conventional methods struggle to build complete shapes and estimate accurate poses due to partial occlusions and sensor noise. They require dense observations to cover all objects, which is challenging to achieve in robotics trajectories. Recent work introduces generative shape priors for object-level mapping from sparse views, but is limited to single-category objects. In this work, we propose a General Object-level Mapping system, GOM, which leverages a 3D diffusion model as shape prior with multi-category support and outputs Neural Radiance Fields (NeRFs) for both texture and geometry for all objects in a scene. GOM includes an effective formulation to guide a pre-trained diffusion model with extra nonlinear constraints from sensor measurements without finetuning. We also develop a probabilistic optimization formulation to fuse multi-view sensor observations and diffusion priors for joint 3D object pose and shape estimation. Our GOM system demonstrates superior multi-category mapping performance from sparse views, and achieves more accurate mapping results compared to state-of-the-art methods on the real-world benchmarks. We will release our code: https://github.com/TRAILab/GeneralObjectMapping.

Figures (12)

• Figure 1: Framework overview. We propose an object-level mapping framework that fuses both multi-view observations and a pre-trained diffusion shape prior model. It generalizes to multi-categories objects, and multiple multi-modalities observations without the need of fine-tuning.
• Figure 2: An illustration of the gradient fields and optimization process. The gradient fields from two sources, a diffusion prior originally for Generation, and multi-view non-linear observation constraints, are effectively fused into a shape and pose optimization formulation for Mapping.
• Figure 3: Prior Effectiveness: Using priors, GOM (Ours) can render higher-quality 3D consistent views and generate 3D meshes with fewer artifacts compared to vMap. Results are based on 10 RGB-D views.
• Figure 4: Mapping Performance: GOM (Ours) outputs 3D object shapes and poses that align with the inputs and further completes the occluded parts. GOM can also generalize to multiple categories. Results are based on 10 RGB-D views.
• Figure 5: Effectiveness of Priors Across 10 Categories on the CO3D Dataset: Toy Truck, Bench, Donut, Broccoli, Toy Train, Apple, Teddy Bear, Hydrant, Book, Toaster, compared to vMap. The results are based on 10 RGB-D views.