IPoD: Implicit Field Learning with Point Diffusion for Generalizable 3D Object Reconstruction from Single RGB-D Images

TL;DR

IPoD tackles the challenge of generalizable 3D object reconstruction from a single RGB-D image by fusing implicit field learning with point diffusion. The method treats query points as a noisy point cloud and denoises them iteratively, yielding adaptive queries that enable both coarse shape recovery and fine-detail precision; a novel self-conditioning mechanism uses implicit predictions to guide diffusion, forming a cooperative loop. Empirical results on CO3D-v2 show state-of-the-art performance, with a 7.8% gain in F-score and a 28.6% improvement in Chamfer distance, and demonstrated generalization on MVImgNet. The work advances practical 3D reconstruction by improving both global geometry and local surface fidelity, with robust generalization across datasets and categories.

Abstract

Generalizable 3D object reconstruction from single-view RGB-D images remains a challenging task, particularly with real-world data. Current state-of-the-art methods develop Transformer-based implicit field learning, necessitating an intensive learning paradigm that requires dense query-supervision uniformly sampled throughout the entire space. We propose a novel approach, IPoD, which harmonizes implicit field learning with point diffusion. This approach treats the query points for implicit field learning as a noisy point cloud for iterative denoising, allowing for their dynamic adaptation to the target object shape. Such adaptive query points harness diffusion learning's capability for coarse shape recovery and also enhances the implicit representation's ability to delineate finer details. Besides, an additional self-conditioning mechanism is designed to use implicit predictions as the guidance of diffusion learning, leading to a cooperative system. Experiments conducted on the CO3D-v2 dataset affirm the superiority of IPoD, achieving 7.8% improvement in F-score and 28.6% in Chamfer distance over existing methods. The generalizability of IPoD is also demonstrated on the MVImgNet dataset. Our project page is at https://yushuang-wu.github.io/IPoD.

Figures (6)

• Figure 1: Our work focuses on the task of generalizable 3D object reconstruction from a single RGB-D image. The proposed method conducts implicit field learning with point diffusion that iteratively denoises a point cloud as adaptive query points for better implicit field learning, which leads to high reconstruction quality on both the global shape and fine details.
• Figure 2: Overview of the proposed method. The network takes a single-view image and a partial point cloud unprojected from the image according to the depth information as the input. A bunch of points are sampled both as a noisy point cloud for point diffusion learning and also as query points for implicit field learning. The proposed self-conditioning mechanism leverages the implicit predictions to reversely assist denoising. The reconstruction result can be obtained via iteratively conducting implicit predicting and denoising concurrently.
• Figure 3: Illustration of the Transformer-based (upper part) and the PVCNN-based (lower part) implementations. $\otimes$ denotes the affine operation. The yellow arrow with double lines indicate the proposed self-conditioning mechanism.
• Figure 4: Visualization on MVImgNet data. Upper: generalization results by Ours2; Lower: the comparison of generalization results before and after fine-tuning on cleaned MVImgNet data.
• Figure 5: Visualization of the denoising process ($t$=$\{$1000, 750, 500, 250, 0$\}$) of our method in inferring. Note we only sample 2k points in each noisy sample for better visualization. The darkness of each point indicates the magnitude of the predicted UDF value (the smaller, the darker) at this position.