FG-Net: A Fast and Accurate Framework for Large-Scale LiDAR Point Cloud Understanding.

Area of Science:

• Computer Vision
• Deep Learning
• Geometric Deep Learning

Background:

• Large-scale point cloud data presents significant challenges for deep learning models.
• Existing methods often rely on voxelization, which can be computationally expensive and lose fine-grained details.
• Real-time processing and high accuracy are critical for practical point cloud understanding applications.

Purpose of the Study:

• To introduce FG-Net, a general deep learning framework for efficient and accurate large-scale point cloud understanding.
• To develop a framework that avoids voxelization while maintaining high performance.
• To demonstrate the robustness and generalization capabilities of the proposed method.

Main Methods:

• A novel noise and outlier filtering technique is employed to preprocess point cloud data.
• A plug-and-play module combines correlated feature mining and deformable convolution for geometric-aware modeling.
• Composite inverse density sampling (IDS) and learning-based operations, along with a feature pyramid residual learning strategy, enhance computational and memory efficiency.

Main Results:

• FG-Net achieves accurate and real-time performance on a single GPU.
• Extensive experiments on eight challenging benchmarks show superior accuracy, speed, and memory efficiency compared to state-of-the-art methods.
• Weakly supervised transfer learning confirms the method's strong generalization capacity.

Conclusions:

• FG-Net provides an effective and efficient solution for large-scale point cloud understanding without voxelization.
• The proposed methods for feature extraction and efficiency improvements significantly advance the field.
• The framework demonstrates potential for real-world applications requiring robust and fast point cloud analysis.