iv-2025-code-release

Getting started

• Install oxford_robotcar_sdk
• Install PVCNN

Creating the required maps

Maps used in this work are point cloud maps that are either 2D slices or 3D maps. For Boreas, lidar maps are first created as one 3D point cloud map using Removert. Then, the 3D map is sliced into 2D maps using the build_2d_lidar_pc_map.py script. Radar maps are directly created as 2D maps using the build_2d_radar_pc_map.py script. For MulRan, 3D point cloud maps are available in the dataset. The 3D maps are sliced into 2D maps using the build_2d_lidar_pc_map.py script. Radar maps are created using the build_2d_radar_pc_map.py script.

All maps created by these scripts need to be chunked into smaller maps. This is done using the chunk_pcd.py script.

Map files should then be moved into dataset and method specific folders. These folders need to be correctly specified in the config files in dataset_cfg.

Training

Training uses the train_residual_ekf_model.py script. We first perform a pretraining using only the scan matching model itself. This is achieved by setting the training_cfg to iv_2025_pretraining. The weights of the pretraining can be loaded into the fine-tuning training config.

Fine Tuning

Fine tuning for the baseline method is done using the train_residual_ekf_model.py script. training_cfg should be set to iv_2025_lbm. For the DCE model, the training_cfg should be set to iv_2025_dce.

Evaluation

Evaluation for all methods is done using the run_localization.py script. Setting experiment to one of

• boreas_radar_icp
• boreas_radar_icp_lidar_map
• mulran_radar_icp
• mulran_radar_icp_lidar_map
• iv_2025_boreas_lbm
• iv_2025_mulran_lbm
• iv_2025_boreas_dce
• iv_2025_mulran_dce

recreates the results from the paper. For MulRan, the sequence that should be used needs to be specified via dataset_cfg, as different parts of the traversed area may require different parts of the map.