kitti-odom-2012-dataloader 0.1.0

Efficient and user-friendly point cloud data loader for the Kitti Odometry 2012 dataset, supporting multiple coordinate systems and numpy compatibility.

KITTI Odometry 2012 Point Cloud Data Loader (English Translation)

This is an English translation of the original 中文版.

Dataset Homepage: KITTI Odometry Benchmark

Project Overview

This project provides an efficient and user-friendly point cloud data loader for the KITTI Odometry 2012 dataset. It supports multiple coordinate system conversions, iteration, slicing, and is suitable for scenarios such as autonomous driving, SLAM, and 3D reconstruction.

Features

• Multiple Coordinate Systems Supported
  • BASE_LINK: Vehicle local coordinate system
  • MAP_KITTI: KITTI official global coordinate system (RDF)
  • MAP_FLU: Common FLU global coordinate system
• Iterator and Slicing: Supports iteration, slicing, and random access
• Numpy Compatible: Can read directly from numpy arrays
• Dataset Length: Supports the len() method

Installation

It is recommended to use uv or pip to install dependencies:

uv pip install -e .

Or use pip directly:

pip install kitti-odom-2012-dataloader

Quick Start

from kitti_odom_2012_dataloader import PointCloudLoader, FrameID

loader = PointCloudLoader("/path/to/kitti/sequences/00", frame_id=FrameID.MAP_FLU)
cld, pose = loader[0]  # Read the first point cloud and its pose

# Iterate over the first 10 frames
for cld, pose in loader[:10]:
    # Process point cloud and pose
    pass

# Get dataset length
print(len(loader))

C++/pybind11 Example

This project supports direct invocation of the Python loader in C++ via pybind11, suitable for integration with C++ projects.
See example/cpp/main.cc for sample code. Ensure the root path points to a valid KITTI dataset directory.

Build and Run Steps

1. Activate Python virtual environment (ensure kitti-odom-2012-dataloader and pybind11 are installed)
2. Modify line 18 in main.cc to set the root variable to your dataset path, e.g., /ws/data/kitti/sequences/00
3. Build the project:

   cd example/cpp
   cmake -S . -B build
   cmake --build build
   

4. Run the example:

   ./build/embed_loader
   

After running, it will output the dataset length, point cloud, and pose shapes.

Main Code Logic

• Start Python interpreter

  #include <pybind11/embed.h>
  namespace py = pybind11;
  py::scoped_interpreter guard{};
  

• Import kitti_odom_2012_dataloader module

  py::module_ kd = py::module_::import("kitti_odom_2012_dataloader");
  py::object PointCloudLoader = kd.attr("PointCloudLoader");
  py::object FrameID = kd.attr("FrameID");
  

• Create PointCloudLoader instance

  const char* root = "/ws/data/kitti/sequences/00";
  py::object loader =
    PointCloudLoader(root, py::arg("frame_id") = FrameID.attr("MAP_FLU"));
  

• Read dataset length

  std::size_t n = loader.attr("__len__")().cast<std::size_t>();
  std::cout << "dataset length = " << n << "\n";
  

• Read point cloud and pose, print shapes

  py::tuple item0 = loader[py::int_(0)];
  py::array_t<float> cld = item0[0].cast<py::array_t<float>>();
  py::array_t<double> pose = item0[1].cast<py::array_t<double>>();
  std::cout << "points shape = (" << cld.shape(0) << ", " << cld.shape(1) << ")\n";
  std::cout << "pose shape   = (" << pose.shape(0) << ", " << pose.shape(1) << ")\n";
  

• Iterate over first 10 frames (optional)

  for (py::size_t i = 0; i < 10 && i < n; ++i) {
    py::tuple it = loader[py::int_(i)];
    // Process point cloud and pose here
  }
  

See example/cpp/main.cc for detailed code.

Coordinate System Description

• BASE_LINK: Vehicle local coordinate system, point cloud transformed by calibration matrix
• MAP_KITTI: KITTI official global coordinate system (RDF), point cloud transformed by global pose
• MAP_FLU: Common FLU global coordinate system, point cloud converted from RDF to FLU

You can switch coordinate systems via the frame_id parameter or attribute:

loader.frame_id = FrameID.MAP_FLU
cld, pose = loader[0]

API Reference

• PointCloudLoader: Main loader class, supports indexing, slicing, iteration
• FrameID: Coordinate system enum type
• get_lidar_files: Get point cloud file list
• read_lidar: Read single frame point cloud
• load_calib_matrix: Read calibration matrix
• load_global_poses: Read global poses
• transform_point_cloud: Point cloud coordinate transformation

Data Format

• Point cloud data shape: [N, 4], representing [x, y, z, intensity]
• Pose matrix shape: [4, 4]

Dependencies

• numpy

Development/testing dependencies (see pyproject.toml):

• open3d
• matplotlib
• pytest
• ipykernel
• ruff

Testing

Unit tests are included. Run:

uv pip install .[dev]
uv run pytest

Changelog

• v0.1.0: Initial version, basic functionality implemented