CMTNet: A hybrid CNN-Mamba-Transformer network for point cloud salient object detection.

Abstract

In recent years, point cloud salient object detection (SOD) has gained increasing attention. However, existing point cloud SOD methods predominantly rely on Farthest Point Sampling (FPS) for downsampling and K-Nearest Neighbors (KNN) or ball query for local feature aggregation, which imposes limitations on both model representation capacity and computational cost. These methods either have a limited receptive field or struggle to balance the performance and efficiency. To address these challenges, our CMTNet employs a serialization-based architecture to achieve efficient downsampling and feature aggregation. Additionally, the proposed CMTNet utilizes 3D sparse convolutions to extract local geometric features rapidly, employs Mamba to model long-range dependencies with linear complexity, and leverages Transformer to capture global context. To adapt Mamba for point cloud tasks, we designed a Point State Space (PSS) block. Furthermore, a Mamba Guided Feature Fusion (MGFF) module is introduced to fuse hierarchical features and enhance long-range dependency modeling. Extensive experiments on the PCSOD dataset demonstrate our approach outperforms current state-of-the-art methods in both effectiveness and efficiency.