DROID Needs Nothing Else: Harnessing Intrinsic Geometric Consistency for Dynamic Gaussian SLAM

Independent motion undermines the geometric consistency of dynamic Gaussian SLAM, biasing camera tracking and embedding transient content into persistent 3D Gaussian maps. Existing optical-flow-based approaches commonly introduce auxiliary perception modules, leaving the robustness inherent in dense correspondence estimation and multi-view optimization underexplored. We present a unified framework that harnesses these native geometric signals to jointly constrain tracking and Gaussian reconstruction. A local linearization analysis motivates complementary reprojection and epipolar residuals to isolate motion unexplained by the static-scene model. Schur-complement curvature yields per-pixel geometric reliability for uncertainty-aware Gaussian depth supervision, while temporal re-examination across keyframe intervals accumulates motion evidence to regulate Gaussian initialization, optimization, and removal. Extensive experiments across diverse datasets demonstrate substantially faster inference and lower computational cost, while matching or surpassing prior-based state-of-the-art methods in tracking and reconstruction accuracy.