Enhancing SplaTAM Performance Through Dynamic Learning Rate Decay and Optimized Keyframe Selection

Abstract

Dense Simultaneous Localization and Mapping (SLAM) is crucial for robotics and augmented reality applications. In this paper, we optimized the SplaTAM using 3D Gaussian representations to achieve high-quality reconstruction using RGB-D cameras. We employ an online tracking and mapping system specifically designed to utilize the underlying Gaussian representations and optimization guided by silhouette rendering. Compared to the original SplaTAM approach, we introduce a dynamic learning rate decay strategy during the camera trajectory tracking phase for tracking. In the mapping phase, we introduce new constraints for selecting keyframes, ensuring that each keyframe contains richer scene information. Finally, we validate our experimental results on the Replica and TUM-RGBD datasets. Our method achieves 14.77% improvement in image rendering performance and 7.04<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> improvement in depth rendering performance over the pre-improvement period, achieving highly competitive performance compared to existing methods.