POSTER
Boyuan Tian, Sushant Kondguli, Dario Seyb, Carl Marshall, Shenlong Wang, Sarita Adve
3D Mapping with Gaussian Splatting (3DGS) has emerged as a promising approach for photorealistic modeling with real-time rendering. Mobile XR use cases require online mapping to promptly reflect content changes, often relying on Simultaneous Localization and Mapping (SLAM) to track camera poses and update map. However, Gaussian-based SLAM systems remain slow due to iterative Gaussian tracking and mapping on dense representations. Our characterization of representative 3DGS-SLAM pipelines reveals three challenges: (1) Gaussian-based pose tracking is expensive and further hinders mapping due to lockstep execution, (2) mapping is dominated by iterative forward rendering and backward gradient aggregation, and (3) bottlenecks are distributed across multiple kernels, making single-point optimization ineffective.
Motivated by the insight that mapping workloads are highly non-uniform across Gaussians and frames, we propose APG-Algorithms, a set of complementary approximations for dominant Gaussian-mapping bottlenecks with tunable speed-quality trade-offs. We further propose APG-Scheduler to jointly control these knobs, maximizing efficiency while constraining quality loss via offline sensitivity profiling and online feedback-driven adaptation. We instantiate these designs in APG-SLAM, built on a representative 3DGS-SLAM system. This instantiation decouples tracking with an off-the-shelf sparse tracker to remove tracking-mapping lockstep dependence, while our technical focus remains approximation-centric acceleration of Gaussian mapping. Evaluations on Replica and TUM-RGBD show that APG-SLAM achieves 2.6x mapping speedup and 2.3x lower power usage, with only a small quality drop in PSNR, SSIM, and LPIPS, capturing both machine-centric and human-perceived fidelity.