KDD-LOAM: Jointly Learned Keypoint Detector and Descriptors Assisted LiDAR Odometry and Mapping

TL;DR

This work addresses robust and efficient 3D point cloud registration by learning a tightly coupled keypoint detector and descriptor (TCKDD) with a self-supervised probabilistic detection loss, implemented on a KPConv-based backbone. The descriptor learning uses a quadruplet contrastive strategy, and the detector learns a saliency-based matchability score that co-adapts to descriptors, enabling robust sparse correspondences. The authors integrate TCKDD into a descriptor-assisted LiDAR odometry and mapping system (KDD-LOAM) featuring a memory-efficient voxel hash map and scan-to-map registration, achieving real-time performance and improved accuracy on KITTI compared to LOAM-family baselines. Extensive experiments on indoor (3DMatch) and outdoor (KITTI) benchmarks demonstrate state-of-the-art registration recall and strong odometry performance, highlighting practical impact for robust SLAM in varied sensing conditions.

Abstract

Sparse keypoint matching based on distinct 3D feature representations can improve the efficiency and robustness of point cloud registration. Existing learning-based 3D descriptors and keypoint detectors are either independent or loosely coupled, so they cannot fully adapt to each other. In this work, we propose a tightly coupled keypoint detector and descriptor (TCKDD) based on a multi-task fully convolutional network with a probabilistic detection loss. In particular, this self-supervised detection loss fully adapts the keypoint detector to any jointly learned descriptors and benefits the self-supervised learning of descriptors. Extensive experiments on both indoor and outdoor datasets show that our TCKDD achieves state-of-the-art performance in point cloud registration. Furthermore, we design a keypoint detector and descriptors-assisted LiDAR odometry and mapping framework (KDD-LOAM), whose real-time odometry relies on keypoint descriptor matching-based RANSAC. The sparse keypoints are further used for efficient scan-to-map registration and mapping. Experiments on KITTI dataset demonstrate that KDD-LOAM significantly surpasses LOAM and shows competitive performance in odometry.

Figures (4)

• Figure 1: The feature maps colored in saliency uncertainty built by our KDD-LOAM on KITTI sequence 07. Sharp corners and edges, distinguishable buildings, pillars, and vehicles are detected as salient regions (red), while flat surfaces, chaotic vegetation, and unstably scanned regions far from the sensor are detected as non-salient regions (blue). It is noteworthy that planar surfaces (most from roads) have been fitted as sparse surfels.
• Figure 2: The network architecture of TCKDD for jointly learning of 3D keypoint detection and description.
• Figure 3: The system overview of KDD-LOAM. We leverage the constant velocity model for scan deskewing and predict the point-wise descriptors and saliency uncertainty through TCKDD. Based on a reliable relative pose guess from RANSAC-based scan-to-scan registration, KDD-LOAM achieves accurate odometry by aligning the deskewed and subsampled scan with a high-resolution yet memory-efficient local map.
• Figure 4: Feature matching recalls on the 3DMatch dataset in relation to inlier distance threshold $\tau_1$ (Left) and inlier ratio threshold $\tau_2$ (Right).