Exploring high-contrast areas context for 3D point cloud segmentation via MLP-driven Discrepancy mechanism

Abstract

Recent advancements in 3D point cloud segmentation, such as PointNext and PointVector, revisit the concise PointNet++ architecture. However, these networks struggle to capture sufficient contextual features in significant high-contrast areas. To address this, we propose a High-contrast Global Context Reasoning (HGCR) module and a Self-discrepancy Attention Encoding (SDAE) block to explore the global and local context in high-contrast regions, respectively. Specifically, HGCR leverages an MLP-driven Discrepancy (MLPD) mechanism and a Mean-pooling function to promote long-range information interactions between high-contrast areas and 3D scene. SDAE expands the degree of freedom of attention weights using an MLP-driven Self-discrepancy (MLPSD) strategy, enabling the extraction of discriminating local context in adjacent high-contrast areas. Finally, we propose a deep network called redPointHC, which follows the architecture of PointNext and PointVector. Our PointHC achieves a mIoU of 74.3% on S3DIS (Area 5), delivering superior performance compared to recent methods, surpassing PointNext by 3.5% and PointVector by 2.0%, while using fewer parameters (22.4M). Moreover, we demonstrate competitive performance with mIoU of 79.8% on S3DIS (6-fold cross-validation), improving upon PointNext by 4.9% and PointVector by 1.4%. Code is available at https://github.com/ShaoyuyuanNEU/PointHC.