DFNET: JOINT GLOBAL CONTEXT MODELING AND LOCAL DETAIL RECOVERY IN POINT CLOUD UPSAMPLING

Abstract

Existing point cloud upsampling methods often encounter several critical limitations, including high computational complexity, excessive numbers of parameters, and suboptimal inference efficiency. These issues primarily stem from overly deep network architectures and redundant feature extraction designs, which hinder their applicability in large-scale or real-time scenarios. To overcome these challenges, we propose DFNet, a novel and efficient dual-branch feature fusion network designed to achieve high-quality point cloud upsampling while significantly reducing computational overhead. The proposed framework consists of two complementary branches that jointly model global structure and local geometric details. Specifically, DFNet incorporates a Spatial Enhancement Attention Module (SAAM) to effectively capture long-range global contextual relationships. By adopting sparse multi-scale self-attention combined with learnable positional encoding, SAAM enhances global feature representation while maintaining a lightweight parameter footprint. In parallel, the Detail Transformation Module (DTM) focuses on recovering fine-grained local details by applying a sparsified self-attention mechanism over local patches. This design efficiently preserves geometric characteristics such as edges, contours, and surface textures, while alleviating the quadratic computational complexity typically associated with standard Transformer-based approaches.