Benchmark for learned camera–LiDAR extrinsic calibration

Learned methods for camera–LiDAR extrinsic calibration report centimetre-level accuracy, but almost always on the dataset and sensor they were trained on. We are building a pre-registered benchmark, on public datasets only, that tests released calibrators on LiDAR types they have not seen (MEMS, FMCW, non-repetitive scanning), using vehicles that carry a familiar and an unfamiliar LiDAR on the same frames. First results on development data show that four released methods lose most or all of their accuracy on any rig other than their training one. To separate the effect of the LiDAR type from that of the training data, the benchmark needs one reference model trained by us on a pool of public driving datasets (KITTI, nuScenes, Argoverse 2, TruckScenes) with a published recipe (CMRNext), unchanged.