We have implemented the pre-training for SemanticKITTI, however the pre-training can be applied to different datasets. Your data should have only the LiDAR point clouds and the scan-wise poses, to aggregate the point clouds. In case poses are not available, you can use kiss-icp to predict the poses. To extract the temporal views, you would need to tune the parameters for the ground segmentation and the hdbscan, to check that the segments correspond to the objects. Note that, in case you face problems using the patchwork method, in this repo they have a pybind version which should make it easier to use, or you could use RANSAC implementation from open3d as done in SegContrast.
In tarl/datasets/dataloader/DataloaderTemplate.py there is a template for implementing the dataloader for you own data, hopefully it has enough comments
and instructions. After implementing your dataloader, be aware to also add it to tarl/datasets/dataloader/datasets.py.
As described in the paper, our pre-training pre-processing consists of two main steps, ground segmentation, and clustering. In this repo the parameters for both method were tunned for SemanticKITTI. In case of pre-training in a different dataset we recommend checking the segments generated in the dataloader and tunning the hyperparameters for those methods to you dataset. Also, even though we use patchwork for ground segmentation and hdbscan for clustering, you can replace those methods with any other method to achieve the segments extraction.