Efficient 3D Point Cloud Feature Learning for Large-Scale Place Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Place recognition using point clouds faces challenges due to environmental appearance changes.
• Current deep learning methods for point cloud retrieval are computationally intensive, limiting their use on resource-constrained hardware.

Purpose of the Study:

• To develop an efficient point cloud learning network (EPC-Net) for place recognition.
• To generate effective global descriptors for point clouds while minimizing computational memory and inference time.

Main Methods:

• Proposed ProxyConv, a lightweight module simplifying edge convolution using adjacency matrices and proxy points for local feature aggregation.
• Designed a grouped VLAD network with grouped fully connected layers to create low-dimensional global descriptors, reducing parameters and maintaining feature discrimination.
• Developed EPC-Net-L, a simplified version using ProxyConv and max pooling, optimized through knowledge distillation.

Main Results:

• EPC-Net and EPC-Net-L achieve competitive performance in place recognition.
• The proposed methods significantly reduce parameters, FLOPs, GPU memory usage, and inference time compared to existing approaches.
• Experiments on multiple datasets validate the efficiency and effectiveness of the developed network.

Conclusions:

• EPC-Net offers an efficient solution for point cloud based place recognition, addressing the limitations of resource-intensive methods.
• The network's design, particularly ProxyConv and grouped VLAD, enables effective global descriptor generation with reduced computational overhead.
• The findings suggest practical applicability in scenarios with limited hardware resources.