iismn/LSV-Loc ? reverse-engineered prompt

Build me the full project for LiDAR to Street View image localization like in the paper.

I want a codebase that can train on paired LiDAR range images and Street View images, learn a shared embedding for place recognition, and then evaluate retrieval plus a PnP based localization step. It should support the main LiDAR setups mentioned in the README, and work with the datasets they list, especially MulRan and the other public or in house formats. Please include the training entry point, the evaluation scripts, dataset loading and preprocessing, configuration files, and the model pieces for the different backbone options like DINOv2, CLIP, and SEA lite.

Make it runnable in a CUDA Linux environment, with support for distributed training if possible. Also add simple visualization for retrieval results and clear instructions for how to prepare the dataset folders and run training or evaluation. If you need to check current library usage or docs online, go ahead and do that.