3DFeat-Net: Weakly Supervised Local 3D Features for Point Cloud Registration

TL;DR

This paper introduces 3DFeat-Net, a weakly supervised method for learning 3D local features for point cloud registration, eliminating the need for manual annotations and achieving state-of-the-art results on outdoor Lidar datasets.

Contribution

It presents a novel weakly supervised approach that learns both 3D feature detection and description directly from GPS/INS tagged data without manual labeling.

Findings

Achieves state-of-the-art performance on outdoor Lidar datasets.

Effectively learns feature correspondences without explicit annotations.

Demonstrates robustness in gravity-aligned point cloud registration.

Abstract

In this paper, we propose the 3DFeat-Net which learns both 3D feature detector and descriptor for point cloud matching using weak supervision. Unlike many existing works, we do not require manual annotation of matching point clusters. Instead, we leverage on alignment and attention mechanisms to learn feature correspondences from GPS/INS tagged 3D point clouds without explicitly specifying them. We create training and benchmark outdoor Lidar datasets, and experiments show that 3DFeat-Net obtains state-of-the-art performance on these gravity-aligned datasets.