About princeton-vl/DROID-SLAM

The official PyTorch implementation of DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras (Teed & Deng, 2021). See the paper page for context.

DROID-SLAM couples a learned dense bundle adjustment layer with recurrent optical flow updates (a descendant of RAFT) to deliver state-of-the-art monocular, stereo, and RGB-D visual SLAM. The repo provides the trained weights, the differentiable BA CUDA kernels, and evaluation scripts for EuRoC, TartanAir, and TUM-RGBD. It is the deep-SLAM baseline that later neural-implicit SLAM systems in the corpus (NICE-SLAM, Co-SLAM, Point-SLAM) benchmark against. Mirrored under BSD-3-Clause — README only.