cilantro: A Lean, Versatile, and Efficient Library for Point Cloud Data Processing

Konstantinos Zampogiannis, Cornelia Fermüller, Yiannis Aloimonos

TL;DR

cilantro is a high-performance, open-source C++ library designed to streamline point cloud data processing. It offers a "Goldilocks" solution for developers: it is more feature-complete than Open3D, significantly faster and more user-friendly than the industry-standard Point Cloud Library (PCL), and supports data in arbitrary dimensions.

What Problem Does This Paper Address?

As robotics, autonomous driving, and AR/VR applications explode, the need to process 3D "point clouds" (collections of data points in space) has become critical. However, developers have long been stuck between two extremes. On one hand, the Point Cloud Library (PCL) is the comprehensive standard but suffers from a steep learning curve, verbose "boilerplate" code, and aging performance due to a lack of modern parallelization. On the other hand, newer libraries like Open3D are easy to use but often lack the breadth of functionality needed for complex research and engineering workflows.

The authors introduce cilantro to bridge this gap, providing a lean, highly optimized, and "expressive" (easy-to-read) C++ framework that handles everything from basic filtering to complex geometric registration.

Methods

The library is built on several modern software engineering principles to ensure both speed and flexibility:

  • Template-Based Design: Most algorithms are templated, allowing them to work with different numerical types (float vs. double) and, crucially, arbitrary dimensions. While most libraries focus strictly on 3D, cilantro can process N-dimensional data.
  • Zero-Copy Data Handling: Using a lightweight wrapper around the Eigen linear algebra library (ConstDataMatrixMap), cilantro can operate on existing memory buffers without needing to copy data, which is a major performance bottleneck in large-scale processing.
  • Modular ICP Framework: The library implements Iterative Closest Point (ICP) using the Curiously Recurring Template Pattern (CRTP). This allows users to easily swap out different components (like point-to-point vs. point-to-plane metrics) while maintaining maximum execution speed.
  • Parallelization: It leverages OpenMP to distribute computationally heavy tasks across multiple CPU cores.

Datasets

To prove cilantro’s mettle, the authors benchmarked it against PCL and Open3D using two well-known datasets:

  1. BundleFusion (apt0 model): A large-scale 3D reconstruction of an apartment consisting of roughly 7.8 million points.
  2. TUM RGBD Dataset (fr1/desk sequence): A standard benchmark for indoor robotic mapping, used here to test normal estimation and frame-to-frame alignment (ICP).

Key Insights and Ideas

  • Expressiveness Matters: One of the most striking features is cilantro’s "method chaining" (e.g., cloud.downsample().estimateNormals().save()). This reduces dozens of lines of PCL code into a single, readable statement, drastically lowering the barrier to entry for new developers.
  • Performance isn't just about the algorithm: By focusing on minimal external dependencies and efficient memory mapping, cilantro shows that "lean" software often outperforms "feature-heavy" giants.
  • Beyond 3D: By supporting arbitrary dimensions, cilantro positions itself as a tool not just for computer vision, but for general-purpose data clustering and spatial reasoning in higher-dimensional feature spaces.

Key Results

The benchmarks show that cilantro consistently outperforms its predecessors:

  • Segmentation Speed: On the 7.8-million-point apt0 model, cilantro performed connected component segmentation in 3.65 seconds, compared to 17.34 seconds for PCL—a nearly 5x speedup.
  • Registration (ICP): In point-to-plane alignment tests, cilantro was 3.82x faster than Open3D and a staggering 14.99x faster than PCL.
  • Normal Estimation: cilantro consistently showed lower running times for surface normal estimation across both k-nearest neighbor (kNN) and radius-based searches compared to both competitors.

Limitations and Future Work

While cilantro is highly optimized for CPUs, the authors acknowledge that it does not yet support GPU acceleration, which could provide even greater speedups for specific tasks. Additionally, while the library covers a wide range of operations, it is still expanding its toolkit for surface reconstruction (turning points into solid meshes) and more advanced non-rigid registration (aligning objects that change shape).

The project is actively maintained and available under the MIT license on GitHub.

Summary updated

Abstract

We introduce cilantro, an open-source C++ library for geometric and general-purpose point cloud data processing. The library provides functionality that covers low-level point cloud operations, spatial reasoning, various methods for point cloud segmentation and generic data clustering, flexible algorithms for robust or local geometric alignment, model fitting, as well as powerful visualization tools. To accommodate all kinds of workflows, cilantro is almost fully templated, and most of its generic algorithms operate in arbitrary data dimension. At the same time, the library is easy to use and highly expressive, promoting a clean and concise coding style. cilantro is highly optimized, has a minimal set of external dependencies, and supports rapid development of performant point cloud processing software in a wide variety of contexts.
BibTeX
@inproceedings{zampogiannis2018cilantro,
  author = {Konstantinos Zampogiannis and Cornelia Fermüller and Yiannis Aloimonos},
  title = {{cilantro: A Lean, Versatile, and Efficient Library for Point Cloud Data Processing}},
  booktitle = {ACM Multimedia},
  year = {2018},
  doi = {10.1145/3240508.3243655},
}

Loading related content...

Figure