We furthermore provide benchmarks related to MOS and static map building along with the dataset. by using our dataset, researchers can evaluate the generalization performance of MOS approaches and static map building approaches against untrained environments and different types of LiDAR sensors.
All the laser scans were deskewed and then saved by utilizing HeLiPR Pointcloud Toolbox. Next, we split the dataset into training, validation, and test sets with ratios of 68%, 16%, and 16%, respectively.
Moving Object Segmentation (MOS)
1. Intersection-over-Union (IoU) for MOS
2. Moving Object Segmentation Performance Across Heterogeneous LiDAR Sensors
L: Livox Avia, A: Aeva Aeries II, O: Ouster OS2-128, and V: Velodyne VLP-16
(a) Mean IoU of MOS approaches trained on the SemanticKITTI dataset
(b) Mean IoU of MOS approaches trained on the SemanticKITTI dataset
(c) Qualitative comparison of MapMOS across all sensors before (top) and after (bottom) training with our dataset.
Static Map Building
1 . Preservation rate (PR), Rejection rate (RR), and F1 score
2. Comparison of static map building approaches for the most crowded frame sequences in our dataset (TBU)
(a) Removert
(b) ERASOR1
(c) ERASOR2
Will be updated.