ScholarWorks@성균관대학교

초록

3D Gaussian Splatting SLAM has emerged as a promising technique for real-time 3D scene representation due to its high rendering quality and compact model size. However, despite its advantages, practical deployment in applications - such as VR/AR systems - remains challenging due to limited processing speed. In particular, the tracking stage presents a critical bottleneck, where the computational overhead significantly affects overall system performance.In this work, we analyze the gradient computation process during the tracking stage of 3D Gaussian Splatting backward rasterization and identify that redundant computations arise from Gaussians associated with negligible loss. We propose a simple yet effective approach that focuses only on significant Gaussians that come from pixels exhibiting high loss and a high object-forming opacity value to guide gradient updates during backward rasterization in tracking. By filtering and computing gradients for these significant Gaussians, our method reduces redundant calculations without substantially sacrificing tracking and rendering quality.This selective tracking approach leads to a substantial reduction in computational cost. Experimental results demonstrate that our proposed method reduces average 90% of computation during tracking, without substantially sacrificing overall tracking and rendering performance. These findings highlight an effective optimization path for accelerating 3D Gaussian Splatting SLAM systems.

키워드