UniReg: A Unified Information Aggregation Framework for Robust Point Cloud Registration

Abstract

Learning discriminative point-wise representations remains the central challenge in scene-level, correspondence-based point cloud registration. Most existing methods process two frames independently during the early stage and introduce cross-frame interaction only at coarsest stages. Such delayed interaction fails to preserve cross-frame fine-grained structural relationships during early hierarchical abstraction, often resulting in ambiguous descriptors under challenging scenarios such as low overlap or repetitive structures. To address this, we propose UniReg, a novel framework that introduces early and continuous cross-frame interaction throughout the entire feature extraction pipeline. At its core, UniReg features a unified backbone composed of three synergistic components: a Uniformer Encoder for fine-grained intra/cross-frame structural relations modeling from the beginning, a Global Aggregation Bottleneck for further aggregating global consistency, and a ConfReg Decoder for dense feature decoding with matchability confidence regularization. These components form a coherent pipeline that preserves cross-frame structural relationships from the earliest stages, enabling UniReg to generate highly discriminative features for robust correspondence establishment. UniReg consistently achieves state-of-the-art performance under standard indoor and outdoor benchmarks, while maintaining competitive computational cost and inference efficiency.