Iterative Feedback Network for Unsupervised Point Cloud Registration

TL;DR

IFNet tackles unsupervised rigid registration of 3D point clouds by introducing an iterative feedback mechanism that allows high-level features to guide low-level representations. It stacks Feedback Registration Blocks (FRBs) and a Feedback Transformer to selectively fuse information across time steps, augmented by a geometry-aware descriptor as a positional embedding. The method defines three losses—global registration, neighborhood consistency, and pseudo consistency—to provide supervision without ground-truth transforms. Experiments across ModelNet40, 7Scenes, ICL-NUIM, and KITTI show state-of-the-art performance with robust, unsupervised registration in the presence of noise and partial overlap.

Abstract

As a fundamental problem in computer vision, point cloud registration aims to seek the optimal transformation for aligning a pair of point clouds. In most existing methods, the information flows are usually forward transferring, thus lacking the guidance from high-level information to low-level information. Besides, excessive high-level information may be overly redundant, and directly using it may conflict with the original low-level information. In this paper, we propose a novel Iterative Feedback Network (IFNet) for unsupervised point cloud registration, in which the representation of low-level features is efficiently enriched by rerouting subsequent high-level features. Specifically, our IFNet is built upon a series of Feedback Registration Block (FRB) modules, with each module responsible for generating the feedforward rigid transformation and feedback high-level features. These FRB modules are cascaded and recurrently unfolded over time. Further, the Feedback Transformer is designed to efficiently select relevant information from feedback high-level features, which is utilized to refine the low-level features. What's more, we incorporate a geometry-awareness descriptor to empower the network for making full use of most geometric information, which leads to more precise registration results. Extensive experiments on various benchmark datasets demonstrate the superior registration performance of our IFNet.

Figures (8)

• Figure 1: The illustrations of the feedback mechanism in our IFNet. Green lines denote the feedback information.
• Figure 2: The overall architecture of IFNet consists of multiple Feedback Registration Block (FRB) modules. Each FRB module generates the feedback information and rigid transformation $\{\mathbf{R}_n^s,\mathbf{t}_n^s\}$, where $n$ represents the number of iterations in the spatial domain and $s$ represents the time step. Additionally, the weight parameters of the FRB modules are shared across time steps.
• Figure 3: The detailed structure of the Feedback Registration Block (FRB).
• Figure 4: The pipeline of the Feedback Transformer.
• Figure 5: The detailed structure of the Matching Matrix Generator.