EPC-Net

EPC-Net generates compact global descriptors from 3D point clouds for large-scale place recognition, optimizing memory and inference time while maintaining retrieval performance.

Key Features:

• ProxyConv Module: ProxyConv aggregates local geometric features using an adjacency matrix and proxy points to simplify edge convolution and reduce GPU memory consumption.
• Grouped VLAD Network: A grouped Vector of Locally Aggregated Descriptors (VLAD) network decomposes high-dimensional vectors into groups via a grouped fully connected layer to reduce parameter count while preserving discriminative power.
• EPC-Net-L Variant: EPC-Net-L comprises two ProxyConv modules followed by max pooling and uses knowledge distillation from the full EPC-Net to produce discriminative global descriptors with lower computational cost.
• Empirical Performance: Validated on the Oxford dataset and three proprietary datasets, reporting reductions in parameters, FLOPs, GPU memory usage, and inference time relative to existing methods.

Scientific Applications:

• Large-scale place recognition: Produces global descriptors for retrieval under seasonal or artificial appearance changes in environments.
• Autonomous navigation: Supports localization and place retrieval in resource-constrained navigation systems subject to environmental variation.
• Augmented reality: Enables efficient point cloud matching for real-time or near-real-time registration in AR scenarios.
• Geographic information systems (GIS): Facilitates efficient 3D point cloud processing and mapping for GIS datasets.

Methodology:

Uses ProxyConv to simplify edge convolution via adjacency matrices and proxy points, employs a grouped VLAD network with a grouped fully connected layer for dimensionality reduction of global descriptors, aggregates descriptors with max pooling in EPC-Net-L, and applies knowledge distillation from the full model to the lightweight variant.