PADLoC: LiDAR-Based Deep Loop Closure Detection and Registration Using Panoptic Attention

A key component of graph-based SLAM systems is the ability to detect loop closures in a trajectory reduce drift accumulated over time from odometry. Most LiDAR-based methods achieve this goal by using only geometric information, disregarding semantics scene. In work, we introduce PADLoC for joint closure detection and registration frameworks. We propose novel transformer-based head point cloud matching registration, leverage panoptic information during training time. particular, loss function that reframes problem as classification task semantic labels graph connectivity assignment instance labels. During inference, does not require annotations, making it more versatile than other methods. Additionally, show two shared heads with their source target inputs swapped increases overall performance enforcing forward-backward consistency. perform extensive evaluations on multiple real-world datasets demonstrating achieves state-of-the-art results. The code our work publicly available at http://padloc.cs.uni-freiburg.de.