A readable, hackable Visual SLAM in 6k lines of C++17

SimpleVisualSLAM is a compact BSD-licensed Visual SLAM system for people who want to read the whole pipeline, modify it, and run experiments without living inside a 50k-line codebase. It supports monocular, RGB-D, stereo, and learned-depth inputs, with loop closing, map persistence, and a ROS2 Jazzy node.

• 6k lines of readable C++17 (vs ORB-SLAM3's 50k)
• Mono / RGB-D / Stereo / DL Depth
• Loop closing with pose graph optimization
• ROS2 Jazzy node included
• BSD-2-Clause license (no GPL contamination)
• 55 unit tests, CI with regression gates

# Clone and build
git clone https://github.com/rsasaki0109/simple_visual_slam.git
cd simple_visual_slam
sudo apt install -y libopencv-dev libeigen3-dev libgoogle-glog-dev libgflags-dev libsuitesparse-dev
cmake -S . -B build -G Ninja -DBUILD_TESTS=ON
cmake --build build -j$(nproc)
ctest --test-dir build

# Run on TUM dataset
./build/run_mono --tum <path-to-tum-sequence> --depth --no-viz

The first configure fetches Ceres Solver, Sophus, DBoW2, and GoogleTest automatically. The full -DUSE_DEPTH_DL=ON build reaches the 55-test configuration.

Stereo is currently available through the EuRoC loader. DL depth requires -DUSE_DEPTH_DL=ON. For the full flag set, run ./build/run_mono --help.

flowchart LR
    A[Main Thread] --> B[Tracking]
    B --> C[LocalMapping]
    C --> D[LoopClosing]

Tracking runs on the main thread. Local mapping and loop closing process keyframes asynchronously and push corrections back into the shared map.

Absolute Trajectory Error (ATE) mean in meters with Sim(3) alignment. These are the current regression-gate snapshot numbers.

On the same TUM head-250 windows and evo_ape settings, stella_vslam currently beats SimpleVisualSLAM in all four measured scenarios. Loop closing did not materially close the gap in this comparison.

The comparison uses the first 250 poses from the provided stella_vslam trajectories, which is the fairest head-window available from the supplied artifacts. If you compare against other systems, keep the dataset window, modality, and evo_ape flags identical. See eval/comparison_protocol.md.

Loop closing looks for data/ORBvoc.txt. If it is missing, the system still runs, but loop closing is disabled.

mkdir -p data
curl -L -o ORBvoc.txt.tar.gz \
    https://github.com/raulmur/ORB_SLAM2/raw/master/Vocabulary/ORBvoc.txt.tar.gz
tar -xzf ORBvoc.txt.tar.gz -C data
rm ORBvoc.txt.tar.gz

Build with -DUSE_DEPTH_DL=ON, then pass either --depth-model <model.onnx> or --metric-depth-model <model.onnx>. Depth Anything v2 works out of the box via ONNX Runtime.

A ROS2 node and launch file live under ros2/. See ros2/README.md for workspace build and launch instructions.

Contributions are welcome. Start with CONTRIBUTING.md, keep changes small and reviewable, and run ctest --test-dir build --output-on-failure before opening a PR. If you touch evaluation logic, keep the regression scripts and benchmark docs in sync.

For papers and reports, please cite the repository (GitHub also reads CITATION.cff). Example BibTeX:

@software{simple_visual_slam,
  title        = {SimpleVisualSLAM},
  year         = {2026},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/rsasaki0109/simple_visual_slam}},
  license      = {BSD-2-Clause},
  version      = {0.1.0},
  note         = {For reproducibility, record the git commit and run ./build/run_mono --version.}
}

This project is licensed under the BSD 2-Clause License.

SimpleVisualSLAM builds on the following open source projects:

• OpenCV for feature extraction, image processing, and visualization
• Ceres Solver for bundle adjustment and nonlinear optimization
• Sophus for Lie group (SE3 / Sim3) operations
• DBoW2 for bag-of-words place recognition
• ONNX Runtime for deep learning inference
• Depth Anything v2 for monocular depth estimation