Mitigating Tracking Failures in Sparse-View Gaussian SLAM via Feature-Guided Initialization

Abstract

Gaussian Splatting has recently emerged as a powerful paradigm for real-time SLAM with high-fidelity scene representation. However, accurate camera tracking remains highly dependent on reliable pose initialization, as photometric optimization requires a good initial estimate to converge. Iterative Closest Point (ICP)-based initialization often fails under challenging conditions such as large viewpoint changes or limited inter-frame overlap, leading to drift and degraded tracking performance. In this work, we address this limitation by introducing a feature-based initialization strategy using SuperPoint and LightGlue, enabling robust correspondence estimation beyond geometric overlap. The estimated pose is obtained via 3D-2D geometric optimization and further refined through photometric alignment, forming a hybrid two-stage tracking pipeline. We evaluate the proposed method on both synthetic and real-world datasets, namely the AriaMultiagent and ReplicaMultiagent datasets. Experimental results demonstrate improved tracking accuracy over the baseline, particularly under sparse-view and large-motion conditions, while maintaining comparable performance in well-conditioned scenarios.