SGSNet: A Lightweight Depth Completion Network Based on Secondary Guidance and Spatial Fusion

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Depth completion is vital for generating dense depth maps from sparse inputs.
• Real-time performance is a key challenge for downstream tasks like SLAM and 3D reconstruction.

Purpose of the Study:

• To propose SGSNet, a lightweight network for efficient depth completion.
• To enhance feature extraction and utilize secondary guidance for accurate depth map generation.

Main Methods:

• Designed an image feature extraction module for multi-scale feature extraction and guidance generation.
• Employed a two-stage guidance mechanism: RGB texture-guided LiDAR feature extraction and sparse depth completion.
• Integrated a lightweight bootstrap module for accelerated network performance.

Main Results:

• SGSNet achieved state-of-the-art accuracy in lightweight depth completion on the KITTI dataset.
• The network demonstrated a 37.5% speed improvement over top published lightweight methods.
• Achieved real-time performance with frame rates up to 30 FPS, suitable for sensor data extraction.

Conclusions:

• SGSNet offers a highly efficient and accurate solution for depth completion.
• The proposed lightweight network meets the speed requirements for complex tasks like SLAM and 3D reconstruction.
• SGSNet represents a significant advancement in real-time depth map generation.