Abstract
Reliable multimodal sensor fusion algorithms require accurate spatiotemporalcalibration. Recently, targetless calibration techniques based on implicitneural representations have proven to provide precise and robust results.Nevertheless, such methods are inherently slow to train given the highcomputational overhead caused by the large number of sampled points requiredfor volume rendering. With the recent introduction of 3D Gaussian Splatting asa faster alternative to implicit representation methods, we propose to leveragethis new rendering approach to achieve faster multi-sensor calibration. Weintroduce 3DGS-Calib, a new calibration method that relies on the speed andrendering accuracy of 3D Gaussian Splatting to achieve multimodalspatiotemporal calibration that is accurate, robust, and with a substantialspeed-up compared to methods relying on implicit neural representations. Wedemonstrate the superiority of our proposal with experimental results onsequences from KITTI-360, a widely used driving dataset.