This project builds a deep-learning-based dense visual SLAM system that performs real-time global optimization of camera poses and 3D geometry. It combines state-of-the-art tracking from DROID-SLAM, monocular depth estimation priors (Metric3D), and differentiable rendering via 3D Gaussian Splatting. The system supports camera intrinsics optimization through DroidCalib and offers alternative branches using MCMC-based densification and 2D Gaussian Splatting for faster convergence.