Related Experiment Video
Updated: Jun 26, 2026
03:31
End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
Published on: December 15, 2023
CMTNet: A hybrid CNN-Mamba-Transformer network for point cloud salient object detection
Langtao Gan1, Hongfa Wen1, Jia Hou2
1School of Communication Engineering, Hangzhou Dianzi University, Hangzhou, 310018, China.
Summary
CMTNet introduces an efficient point cloud salient object detection (SOD) method using serialization and Mamba for improved performance. This approach enhances both effectiveness and computational efficiency in 3D data analysis.
Area of Science:
- Computer Vision
- 3D Data Processing
- Machine Learning
Background:
- Point cloud salient object detection (SOD) is gaining attention but faces challenges with existing methods relying on Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN) or ball query.
- Current methods exhibit limitations in representation capacity and computational cost, with restricted receptive fields or poor performance-efficiency balance.
Purpose of the Study:
- To develop a novel point cloud SOD method, CMTNet, that overcomes the limitations of existing approaches.
- To achieve efficient downsampling and feature aggregation using a serialization-based architecture.
- To enhance model representation capacity and computational efficiency for salient object detection in point clouds.
Main Methods:
- CMTNet employs a serialization-based architecture for efficient downsampling and feature aggregation.
- Utilizes 3D sparse convolutions for rapid local geometric feature extraction.
- Incorporates Mamba with a novel Point State Space (PSS) block for linear-complexity long-range dependency modeling.
- Leverages Transformer for global context capture and a Mamba Guided Feature Fusion (MGFF) module for hierarchical feature fusion.
Main Results:
- CMTNet demonstrates superior performance compared to state-of-the-art methods on the PCSOD dataset.
- The proposed method achieves a better balance between effectiveness and computational efficiency.
- The serialization-based architecture and Mamba integration significantly improve long-range dependency modeling and feature fusion.
Conclusions:
- CMTNet presents an effective and efficient solution for point cloud salient object detection.
- The integration of Mamba and Transformer architectures offers a promising direction for future 3D vision tasks.
- The developed Point State Space (PSS) block and Mamba Guided Feature Fusion (MGFF) module are key innovations for enhancing point cloud analysis.