// the find
rmurai0610/MASt3R-SLAM
[CVPR 2025] MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors
MASt3R-SLAM is a dense monocular/RGB-D SLAM system built on top of NAVER's MASt3R foundation model, which gives it strong 3D reconstruction priors without needing classical feature descriptors. It runs in real-time on a modern GPU and outputs a dense point cloud map alongside camera trajectories. Aimed at CV researchers and roboticists who want a SLAM baseline that doesn't fall apart on textureless surfaces.
The core idea is genuinely clever: MASt3R's metric depth predictions act as a prior that sidesteps the bootstrapping problem traditional monocular SLAM has in low-texture scenes. The calibration-free mode is practically useful — most real deployments involve cameras whose exact intrinsics you don't fully trust. Custom CUDA kernels for Gauss-Newton pose updates and matching keep it real-time rather than being a research demo that takes 10x slower-than-realtime. Evaluation scripts for TUM, EuRoC, 7-Scenes, and ETH3D are all wired up, so comparing against it in a paper is actually straightforward.
The dependency chain is brutal: you're installing PyTorch, three nested git submodules (mast3r → dust3r → croco), CUDA extensions, and a custom OpenGL visualizer just to run it. The WSL workaround of disabling multiprocessing entirely is a code smell — shared memory issues on Windows suggest the process architecture isn't portable. No CPU fallback path at all, which means an RTX-class GPU isn't optional. Last commit was November 2025 with the repo essentially frozen post-paper, so bug reports are piling up without active maintenance.