OSM-BKI: Bayesian Kernel Inference for LiDAR Semantic Segmentation

Bayesian Kernel Inference (BKI) for semantic label refinement of LiDAR point clouds using OpenStreetMap (OSM) priors.

This project features a Continuous Mapping Engine (continuous_bki) as its core backend, supporting both single-scan refinement and stateful semantic mapping over time.

---

Project Structure

OSM-BKI/
├── README.md                 # This file
├── configs/                  # Dataset configurations
│   ├── mcd_config.yaml       # MCD (KTH campus) – direct paths or dataset_root_path+sequence
│   └── kitti*.yaml           # KITTI / KITTI360 configs
│
├── example_data/             # Sample data
│   ├── mcd/                  # MCD format (KTH)
│   │   ├── kth_day_06/       # LiDAR, labels, poses
│   │   ├── kth.osm           # OSM map
│   │   └── hhs_calib.yaml    # Calibration
│   └── kitti360/             # KITTI360 format
│
├── python/                   # Python package, scripts, benchmarks
│   ├── README.md             # Python setup & usage
│   ├── run_osmbki.sh         # Run continuous map on example data
│   ├── run_benchmarks.sh     # Run all benchmarks
│   ├── scripts/              # continuous_map_train_test, visualize_osm_xml, bki_tools, etc.
│   └── benchmarks/           # Throughput, calibration, OSM modes, etc.
│
└── cpp/osm_bki/              # C++ library
    ├── README.md             # C++ build & examples
    └── src/, include/        # Core engine

---

Quick Start

1. Build the Python extension

cd python
pip install -r requirements.txt
pip install -e .

2. Run on example data

From the repo root:

./python/run_osmbki.sh

Trains a continuous BKI map on example_data/mcd/kth_day_06 and evaluates on held-out scans. Output: output.bki.

3. Run benchmarks

./python/run_benchmarks.sh

---

Configuration

Paths are specified via direct keys in the config (relative to the config file) or via legacy dataset_root_path + sequence:

• Direct paths: lidar_dir, label_dir, pose_path, calibration_path, osm_file
• Legacy: dataset_root_path + sequence – paths derived under dataset_root_path/sequence/

See configs/mcd_config.yaml and python/README.md for details.

---

Python API

import osm_bki_cpp
import numpy as np

# Initialize BKI engine
bki = osm_bki_cpp.PyContinuousBKI(
    osm_path="example_data/mcd/kth.osm",
    config_path="configs/mcd_config.yaml",
    resolution=0.5,
    l_scale=1.0
)

# Load data (numpy arrays)
points = np.fromfile("scan.bin", dtype=np.float32).reshape(-1, 4)[:, :3]
labels = np.fromfile("labels.bin", dtype=np.uint32)

# Update map and infer refined labels
bki.update(labels, points)
refined = bki.infer(points)

# Save map state
bki.save("map.bki")

---

Features

• Unified backend: Single C++ engine (ContinuousBKI) for all operations
• Voxel-based: Sparse voxel hashing for O(1) lookups
• Stateful: Accumulates semantic evidence across thousands of scans
• OSM priors: XML or binary OSM geometries for roads, buildings, vegetation, etc.
• Multi-threaded: OpenMP support

---

Requirements

• Python 3.7+
• NumPy, pybind11, PyYAML, Open3D (see python/requirements.txt)
• C++17 compiler with OpenMP

---

Documentation

• python/README.md – Python setup, scripts, benchmarks
• cpp/osm_bki/README.md – C++ build and examples

---

Citation

If you use this work, please cite:

@software{osm_bki,
  title={OSM-BKI: Bayesian Kernel Inference for LiDAR Semantic Segmentation with OpenStreetMap Priors},
  author={OSM-BKI Team},
  year={2026},
  url={https://github.com/yourrepo/OSM-BKI}
}

---

License

MIT License