Fast and Precise Multimodal Spatiotemporal Calibration via Periodic-Activated 2D Gaussian Splatting

Hongbeen Park, Minjeong Park, Sunpil Kim, Jinkyu Kim, Jung Hyun Lee

Keywords: Multimodal SpatioTemporal Calibration, Periodic Activation, Gaussian Splatting

Abstract: Accurate spatiotemporal calibration between heterogeneous sensors such as cameras and LiDAR is essential for robust performance in tasks like localization, mapping, and object detection. Traditional calibration methods rely on physical targets and manual procedures, limiting scalability and applicability in dynamic or real-time environments. Recent advances in targetless and scene reconstruction-based calibration have improved spatial alignment but often neglect temporal synchronization and are computationally intensive. In this work, we propose PACS (Periodic Activation with 2D Gaussian Splatting for Multimodal SpatioTemporal Calibration), a novel and efficient calibration framework that jointly estimates spatial and temporal alignment across sensors. PACS uses a sine-activated multilayer perceptron (MLP) to effectively capture high-frequency scene details, enhancing convergence and representational power. Leveraging an anchor-based decoding scheme, our method significantly accelerates training while maintaining robust scene reconstruction. Furthermore, we employ 2D Gaussian splatting to render scenes with improved alignment between visual and LiDAR-based geometry. Experimental results demonstrate that PACS achieves accurate, efficient, and robust calibration across multimodal sensor configurations.

Submission Number: 39