// the find
gaoxiang12/ORBSLAM2_with_pointcloud_map
A fork of ORB_SLAM2 that bolts on a live point cloud map built from RGB-D keyframes, so you get a visual/dense map alongside the usual sparse feature-point SLAM output. Useful for robotics folks who want a quick dense map without switching to a different SLAM stack entirely.
The core contribution is narrow and understandable: it hooks into the existing RGB-D pipeline and accumulates a point cloud as keyframes come in, rather than reinventing SLAM. It inherits ORB_SLAM2's actual tracking/loop-closing, which is a proven, well-cited algorithm, so you're not trusting an unknown implementation for the hard part.
The repo has compiled build artifacts and CMake cache files checked into git (object files, binaries, CMakeCache.txt) under the ROS example directory, which bloats the repo and is a sign nobody bothered with a .gitignore. The README gives zero detail on how the point cloud module actually works (voxel filtering? octree? raw accumulation?) or its performance/memory cost, which matters a lot for anything beyond a toy dataset. Dependency chain is heavy and brittle: a hand-patched g2o fork, Pangolin, DBoW2, and the separately-downloaded ORB vocabulary file, none of which are pinned to versions, so builds that worked in 2022 are a gamble today. No activity since August 2022 and no tests, so any breakage on a modern Ubuntu/OpenCV/PCL toolchain is on you to fix.