Open3D C++ API

Find the right Open3D C++ symbol, verify it against the installed headers, and write code that follows current Open3D direction: tensor API first, Filament visualization first.

Applies to Open3D v0.20.0. This skill covers the public API for applications that use Open3D, not Open3D's own internals.

When to Use

• Writing or reviewing C++ that includes open3d/... headers
• Choosing between open3d::geometry (legacy) and open3d::t::geometry (tensor)
• Looking up a signature, dtype/device contract, or class hierarchy
• Linking Open3D from a CMake project

API Direction (decide this first)

Fall back to the legacy API only when there is no tensor equivalent (see the legacy-only list in references/api-map.md). Convert at the boundary with FromLegacy() / ToLegacy() and note the reason in a comment.

Never add a silent CPU fallback for a CUDA/SYCL path, and never move data between devices implicitly.

Get Open3D

Do not build Open3D from source unless you have to — it takes a long time. Download a prebuilt binary package instead:

• Releases: https://github.com/isl-org/Open3D/releases — open3d-devel-*.tar.xz for Linux, macOS, and Windows, in CPU and CUDA variants
• Extract it and point CMake at it with -DOpen3D_ROOT=/path/to/open3d-devel-...

If you genuinely must build from source, follow https://www.open3d.org/docs/latest/compilation.html rather than improvising.

This skill ships inside that package at share/Open3D/agent_skills/open3d-cpp (bin/Open3D/agent_skills/open3d-cpp on Windows). Copy it into your project's .github/skills/ to make it available to an agent.

Discovery Procedure

The installed headers are authoritative — they match the binary you are linking against. Doxygen HTML is generally not installed locally; use the website.

1. Grep the installed headers — start here

O3D=/path/to/open3d-devel-.../include        # or /usr/local/include
grep -rn "class PointCloud" "$O3D/open3d/t/geometry/"
grep -rn "FromLegacy\|ToLegacy" "$O3D/open3d/t/geometry/PointCloud.h"
grep -rn "MultiScaleICP" "$O3D/open3d/t/pipelines/registration/"

Declarations carry Doxygen \brief, \param, and \return comments — that is the primary C++ reference. rg works equally well. The umbrella header open3d/Open3D.h lists every public sub-header and is the quickest way to see the whole surface.

To find a class when you do not know its header:

grep -rln "class RaycastingScene" "$O3D/open3d/"

2. Online API reference and tutorials

• C++ API reference: https://www.open3d.org/docs/latest/cpp_api/ — use the site's search box
• Tutorials and conceptual docs: https://www.open3d.org/docs/latest/
• Source, issues, and the full example tree: https://github.com/isl-org/Open3D — use GitHub code search, e.g. repo:isl-org/Open3D MultiScaleICP path:examples/cpp

3. Examples on GitHub

C++ examples live in examples/cpp/. Useful entry points: PointCloud.cpp, TriangleMesh.cpp, TICP.cpp, RegistrationRANSAC.cpp, TIntegrateRGBD.cpp, Draw.cpp, OffscreenRendering.cpp, MultipleWindows.cpp, Visualizer.cpp.

If you also have Open3D's Python wheel installed, open3d example --list gives runnable Python equivalents that mirror the C++ API one-to-one (snake_case instead of PascalCase).

Linking Open3D

cmake_minimum_required(VERSION 3.24)
project(MyApp LANGUAGES CXX)

find_package(Open3D REQUIRED)

add_executable(MyApp main.cpp)
target_link_libraries(MyApp PRIVATE Open3D::Open3D)

cmake -S . -B build -DOpen3D_ROOT=/path/to/open3d-devel-...
cmake --build build

Open3D requires a C++17-capable compiler and CMake 3.24+. Working templates: examples/cmake/open3d-cmake-find-package (prebuilt package — recommended) and examples/cmake/open3d-cmake-external-project (builds Open3D alongside your project — slow).

pkg-config also works on Linux/macOS with shared libraries, but CMake is strongly preferred because it handles the optional backends correctly:

export PKG_CONFIG_PATH="$PKG_CONFIG_PATH:<install>/lib/pkgconfig"
c++ main.cpp -o app $(pkg-config --cflags --libs Open3D)   # libs must follow sources

Full details: https://www.open3d.org/docs/latest/cpp_project.html

Snippets

Adapted from the official examples; see each link for the full program.

Read, process, write a point cloud

#include "open3d/Open3D.h"
using namespace open3d;

t::geometry::PointCloud pcd;
t::io::ReadPointCloud("input.ply", pcd);

auto down = pcd.VoxelDownSample(0.05);
down.EstimateNormals(30, 0.1);
auto [clean, mask] = down.RemoveStatisticalOutliers(20, 2.0);

t::io::WritePointCloud("output.ply", clean);
utility::LogInfo("{} -> {} points", pcd.GetPointPositions().GetLength(),
                 clean.GetPointPositions().GetLength());

Tensors and devices

core::Device device = core::cuda::IsAvailable() ? core::Device("CUDA:0")
                                                : core::Device("CPU:0");
core::Tensor points = core::Tensor::Zeros({100, 3}, core::Float32, device);

// From existing memory (copies into an Open3D-owned blob)
std::vector<float> raw{0, 0, 0, 1, 0, 0, 0, 1, 0};
core::Tensor from_raw(raw, {3, 3}, core::Float32, device);

t::geometry::PointCloud pcd(points);
pcd.SetPointAttr("colors", core::Tensor::Ones({100, 3}, core::Float32, device));

auto on_cpu = pcd.To(core::Device("CPU:0"));   // explicit; nothing moves implicitly

C++ dtype constants are capitalized (core::Float32, core::Int64); the Python equivalents are lowercase.

Point-to-plane multi-scale ICP

using namespace open3d::t::pipelines::registration;

std::vector<double> voxel_sizes{0.05, 0.025, 0.0125};
std::vector<double> max_correspondence_distances{0.1, 0.05, 0.025};
std::vector<ICPConvergenceCriteria> criteria;
criteria.emplace_back(1e-6, 1e-6, 30);
criteria.emplace_back(1e-6, 1e-6, 15);
criteria.emplace_back(1e-6, 1e-6, 10);

source.EstimateNormals();
target.EstimateNormals();

auto result = MultiScaleICP(
        source, target, voxel_sizes, criteria, max_correspondence_distances,
        core::Tensor::Eye(4, core::Float64, core::Device("CPU:0")),
        TransformationEstimationPointToPlane());
utility::LogInfo("fitness {} rmse {}", result.fitness_, result.inlier_rmse_);

result.transformation_ is always Float64 on CPU:0. Source: TICP.cpp.

Ray casting and signed distance

auto mesh = t::geometry::TriangleMesh::CreateSphere(1.0);
t::geometry::RaycastingScene scene;            // or scene(0, core::Device("SYCL:0"))
scene.AddTriangles(mesh);

auto rays = t::geometry::RaycastingScene::CreateRaysPinhole(
        60.0, core::Tensor::Init<float>({0, 0, 0}),   // center
        core::Tensor::Init<float>({0, 0, 3}),         // eye
        core::Tensor::Init<float>({0, 1, 0}),         // up
        640, 480);
auto result = scene.CastRays(rays);            // map: "t_hit", "geometry_ids", ...

auto query = core::Tensor::Init<float>({{0, 0, 0}, {2, 0, 0}});
auto sdf = scene.ComputeSignedDistance(query);

TSDF integration

t::geometry::VoxelBlockGrid vbg(
        {"tsdf", "weight", "color"},
        {core::Float32, core::Float32, core::Float32},
        {{1}, {1}, {3}}, 3.0f / 512, 16, 50000, device);

for (size_t i = 0; i < depth_files.size(); ++i) {
    auto depth = t::io::CreateImageFromFile(depth_files[i])->To(device);
    auto color = t::io::CreateImageFromFile(color_files[i])->To(device);
    auto blocks = vbg.GetUniqueBlockCoordinates(*depth, intrinsic, extrinsics[i],
                                                1000.0f, 3.0f);
    vbg.Integrate(blocks, *depth, *color, intrinsic, intrinsic, extrinsics[i],
                  1000.0f, 3.0f);
}
auto pcd = vbg.ExtractPointCloud();            // weight threshold defaults to 3.0
auto mesh = vbg.ExtractTriangleMesh();

Source: TIntegrateRGBD.cpp.

Nearest neighbor search

core::Tensor points = core::Tensor::Init<float>({{0, 0, 0}, {1, 0, 0}, {0, 1, 0}});
core::nns::NearestNeighborSearch nns(points);

nns.KnnIndex();
auto [indices, distances2] = nns.KnnSearch(points, 2);

nns.FixedRadiusIndex(0.5);
auto [r_idx, r_dist2, r_splits] = nns.FixedRadiusSearch(points, 0.5);

Visualization

auto mesh = std::make_shared<geometry::TriangleMesh>();
io::ReadTriangleMesh("mesh.ply", *mesh);
mesh->ComputeVertexNormals();
visualization::Draw({mesh});                    // one-liner, Filament-backed

Named objects and per-object visibility:

visualization::Draw({visualization::DrawObject("source", source),
                     visualization::DrawObject("target", target, false)});

Offscreen rendering:

using namespace open3d::visualization;

auto &app = gui::Application::GetInstance();
app.Initialize();

auto *renderer = new rendering::FilamentRenderer(
        rendering::EngineInstance::GetInstance(), 640, 480,
        rendering::EngineInstance::GetResourceManager());
auto *scene = new rendering::Open3DScene(*renderer);

rendering::MaterialRecord material;
material.shader = "defaultLit";
scene->AddGeometry("mesh", mesh.get(), material);

auto image = app.RenderToImage(*renderer, scene->GetView(), scene->GetScene(),
                               640, 480);
io::WriteImage("render.png", *image);

Sources: Draw.cpp, OffscreenRendering.cpp.

Conventions and Gotchas

• Naming: C++ is PascalCase where Python is snake_case (VoxelDownSample ↔ voxel_down_sample). Getters are explicit in C++: GetPointPositions() / SetPointAttr() versus Python's pcd.point["positions"].
• Tensors are shape + strides + blob + dtype + device. Copies are shallow by default; use Clone() for a deep copy and Contiguous() when a kernel needs it. Dtype constants are core::Float32, core::Float64, core::Int64, core::Bool.
• Device is explicit. All inputs to a tensor pipeline must be on the same device; To(device) each one.
• Indexing uses core::TensorKey::Index/Slice/IndexTensor with core::None for an open slice bound, following NumPy semantics.
• utility::LogError is [[noreturn]] and throws std::runtime_error — it is the error mechanism, not just a print. Also LogWarning, LogInfo, LogDebug; set the level with utility::SetVerbosityLevel.
• RaycastingScene supports CPU and SYCL devices, not CUDA. Select with the constructor's device argument. (The class comment in RaycastingScene.h still says CPU-only; the binding docs and implementation are correct.)
• VoxelBlockGrid::ExtractPointCloud/ExtractTriangleMesh take a weight threshold defaulting to 3.0; a voxel needs that many observations to appear.
• Registration transformations are always Float64 on CPU:0.
• Legacy and tensor types are distinct — geometry::PointCloud and t::geometry::PointCloud do not interconvert implicitly.
• CUDA/SYCL availability is a property of the binary you downloaded. Check with core::cuda::IsAvailable() / core::sycl::IsAvailable() at runtime.

Reference Files

• references/api-map.md — namespace map with headers, per-class method inventories, tensor↔legacy mapping, legacy-only and tensor-only lists