ALT Our method combines moving object segmentation (b) and instance segmentation (c) to solve the panoptic task of moving instance segmentation from sparse radar point clouds (a). The reference image in (d) illustrates the scene and includes privacy-preserving colored masks. In the image of the point cloud (c), each color represents a different instance of moving objects (static is gray). The colors in (c) and (d) correspond if the object is visible.

ALT Computation of the TOPS descriptor from slices of an aligned object point cloud, P^ , using the slicing-based filtration. The figure shows a pictorial representation of the filtration ( C0…C∞ ) constructed from one of the slices of P^ . The origin and termination points of the slice (shown in black and blue, respectively) enable tracking of the birth and death of connected components in the filtration. Every connected component in the PD corresponds to a “strip” in the 2-D slice. The PIs obtained from the PD of every slice are vectorized and stacked to form the TOPS descriptor.

ALT Point cloud map of an actual shipwreck collected in Wiarton, ON, Canada, where color represents relative depth. The top map is generated using the state estimate produced by a COTS Doppler velocity log-aided inertial measurement system (DVL-INS). The bottom map is generated using the proposed method, which conditions the DVL-INS estimate on loop-closure measurements without access to the raw DVL-INS sensor measurements. Note the improvement in map self-consistency when loop-closure measurements are included.