Open3D Python API Map

Open3D v0.20.0.

This file does not list class members — dir() and help() already do that, better, and always in sync with your installed build:

python -c "import open3d as o3d; print([n for n in dir(o3d.t.geometry.PointCloud) if not n.startswith('_')])"
python -c "import open3d as o3d; help(o3d.t.geometry.PointCloud.voxel_down_sample)"

What follows is what introspection cannot tell you: which namespace owns a task, which of ~60 similarly-named methods is the right one, what is current versus legacy, and which behaviors will surprise you.

Namespace Routing

You want to…	Namespace
Numeric arrays, devices, linear algebra	open3d.core
Neighbor search	open3d.core.nns
Point clouds, meshes, images, bounding volumes, ray casting, TSDF	open3d.t.geometry
Read/write geometry and images; RealSense	open3d.t.io
ICP, feature matching, registration	open3d.t.pipelines.registration
RGB-D odometry	open3d.t.pipelines.odometry
Dense SLAM	open3d.t.pipelines.slam
Non-rigid refinement of fragment poses	open3d.t.pipelines.slac
Show something on screen	open3d.visualization (draw, O3DVisualizer)
Build a custom 3D app	open3d.visualization.gui + .rendering
Sample data	open3d.data
Legacy Eigen equivalents (fallback only)	open3d.geometry, .io, .pipelines, .camera

Task → API

Point clouds

Task	Call
Load / save	t.io.read_point_cloud, t.io.write_point_cloud
Reduce point count	voxel_down_sample (uniform grid, usual choice), uniform_down_sample (every k-th), random_down_sample, farthest_point_down_sample (best coverage, slowest)
Normals	estimate_normals, then one of orient_normals_to_align_with_direction / orient_normals_towards_camera_location / orient_normals_consistent_tangent_plane
Denoise	remove_statistical_outliers (density-adaptive), remove_radius_outliers (fixed radius), remove_non_finite_points
Take a subset	select_by_index, select_by_mask, crop
Find planes / clusters	segment_plane (one plane, RANSAC), cluster_dbscan (arbitrary clusters)
Smooth	smooth_mls, smooth_bilateral, smooth_laplacian, smooth_taubin
Move / colorize	transform, translate, rotate, scale, paint_uniform_color
From RGB-D	create_from_depth_image, create_from_rgbd_image (static)
Back to 2D	project_to_depth_image, project_to_rgbd_image
Compare two clouds	compute_metrics (Chamfer, Hausdorff, F-score)

Meshes

Task	Call
Load / save	t.io.read_triangle_mesh, t.io.write_triangle_mesh
Primitives	create_box, create_sphere, create_cylinder, create_cone, create_torus, create_arrow, create_coordinate_frame, create_text (static)
Normals	compute_vertex_normals (smooth shading), compute_triangle_normals (flat)
Reduce triangles	simplify_quadric_decimation(target_reduction=…)
Repair	fill_holes, remove_non_manifold_edges, remove_unreferenced_vertices
CSG	boolean_union, boolean_intersection, boolean_difference
Cut	clip_plane (keep a half-space), slice_plane (cross-section curves)
Mesh → points	sample_points_uniformly
UVs and textures	compute_uvatlas, bake_vertex_attr_textures, bake_triangle_attr_textures, project_images_to_albedo
Surface from a point cloud	legacy only — geometry.TriangleMesh.create_from_point_cloud_poisson / _alpha_shape / _ball_pivoting

Registration

Task	Call
Refine a roughly-aligned pair	multi_scale_icp (preferred) or icp
Pick the error metric	TransformationEstimationPointToPlane (needs normals, converges fastest), …PointToPoint, …ForColoredICP (needs colors), …Symmetric, …ForDopplerICP
Downweight outliers	pass a robust_kernel to the estimation method
Coarse / global alignment	compute_fpfh_feature + correspondences_from_features; full RANSAC/FGR is legacy only (pipelines.registration.registration_ransac_based_on_feature_matching, registration_fgr_based_on_feature_matching)
Score an alignment	evaluate_registration
Covariance for a pose graph	get_information_matrix
Optimize many fragments jointly	legacy only — pipelines.registration.PoseGraph + global_optimization

Reconstruction

Task	Call
Fuse RGB-D frames into a volume	t.geometry.VoxelBlockGrid — compute_unique_block_coordinates, then integrate
Get the surface out	extract_point_cloud, extract_triangle_mesh
Frame-to-frame camera pose	t.pipelines.odometry.rgbd_odometry_multi_scale
Online dense SLAM	t.pipelines.slam.Model + Frame
Refine fragment poses	t.pipelines.slac.run_slac_optimizer_for_fragments

Queries and search

Task	Call
k nearest neighbors	core.nns.NearestNeighborSearch — knn_index() then knn_search
All neighbors within a radius	fixed_radius_index(r) then fixed_radius_search
A different radius per query	multi_radius_index / multi_radius_search
k neighbors capped by a radius	hybrid_index / hybrid_search
Ray/mesh intersection	t.geometry.RaycastingScene.cast_rays
Distance to a surface	compute_distance (unsigned), compute_signed_distance (signed)
Inside/outside test	compute_occupancy
Closest point on a surface	compute_closest_points
Visibility / counting hits	test_occlusions, count_intersections, list_intersections

KDTreeFlann is legacy — always use core.nns instead.

Visualization

Task	Call
Just look at it	visualization.draw(geometry)
Several named, toggleable objects	visualization.draw([{"name": …, "geometry": …}, …]) or O3DVisualizer
An image file, no window	rendering.OffscreenRenderer → render_to_image
Custom app with widgets	gui.Application + gui.Window + gui.SceneWidget + rendering.Open3DScene
Control appearance	rendering.MaterialRecord — set .shader to defaultLit, defaultUnlit, normals, depth, or unlitLine
Inline in a notebook	visualization.draw_plotly
Log 3D data over training steps	visualization.tensorboard_plugin (needs tensorboard)

Tensor ↔ Legacy

Same class name, different namespace: PointCloud, TriangleMesh, LineSet, Image, RGBDImage, AxisAlignedBoundingBox, OrientedBoundingBox, and every read_* / write_* IO function (o3d.io.X → o3d.t.io.X).

Renamed or restructured:

Legacy	Tensor
pipelines.registration.registration_icp	t.pipelines.registration.icp / multi_scale_icp
get_information_matrix_from_point_clouds	t.pipelines.registration.get_information_matrix
pipelines.odometry.compute_rgbd_odometry	t.pipelines.odometry.rgbd_odometry_multi_scale
pipelines.integration.ScalableTSDFVolume	t.geometry.VoxelBlockGrid (block-sparse; not drop-in)
geometry.KDTreeFlann + KDTreeSearchParam*	core.nns.NearestNeighborSearch — always migrate
pcd.points / .colors / .normals	pcd.point["positions" / "colors" / "normals"]
visualization.draw_geometries / Visualizer	visualization.draw / O3DVisualizer / gui + rendering

Legacy-only — a fallback here is justified

Octree · TetraMesh · HalfEdgeTriangleMesh · geometry.keypoint · TSDF volumes (pipelines.integration.*) · pose graphs and global_optimization · RANSAC and FGR global registration · pipelines.color_map.* · camera.* · surface reconstruction from point clouds (Poisson / alpha shape / ball pivoting) · visualizer editing and vertex-selection subclasses · detect_planar_patches · compute_point_cloud_distance · voxel_down_sample_and_trace

Tensor-only — no legacy equivalent

core.Tensor and explicit device placement · core.HashMap / HashSet · core.nns.NearestNeighborSearch · RaycastingScene · VoxelBlockGrid · custom per-point attributes via TensorMap · Chamfer/Hausdorff/F-score metrics · SLAC and SLAM · Doppler ICP

Non-Obvious Behavior

Things that cost debugging time and are not visible in a signature.

API	What to know
dtype constants	Lowercase in Python (o3d.core.float32); Float32 is C++ only
linear algebra	There is no o3d.core.linalg — matmul, inv, det, svd, solve, lstsq, lu are Tensor methods
from_legacy	Defaults to Float32 even though legacy geometry stores float64
tensor geometry attributes	pcd.point is a TensorMap with no .keys() — use .primary_key or dir()
multi_scale_icp	voxel_sizes and max_correspondence_distances must be o3d.utility.DoubleVector, not lists
any registration result	.transformation is always Float64 on CPU:0, whatever device the inputs were on
extract_point_cloud / extract_triangle_mesh	weight_threshold=3.0 by default — a voxel needs ≥3 observations, so few-frame integrations come back empty
RaycastingScene	CPU and SYCL only, never CUDA. Device is a constructor argument
compute_signed_distance	The sign is only meaningful for watertight meshes
fixed_radius_search	Ragged output: values plus row-split offsets, not a rectangular array
estimate_normals	Orientation is arbitrary; follow with an orient_normals_* call
index tensors	Must be Int64
Tensor.numpy()	The tensor must already be on CPU
from_dlpack	Shares memory; a to_dlpack capsule can be consumed only once
o3d.data	Downloads on first use into ~/open3d_data; requires network access
gui	Application.instance.initialize() before any window; main thread only
optional modules	webrtc_server, ml.torch, ml.tf, tensorboard_plugin may be absent — guard with importlib.util.find_spec

Sample Data

o3d.data classes download on construction; read .path, .paths, or type-specific attributes (.depth_paths, .color_paths, .camera_intrinsic_path).

Need	Use
A point cloud	PCDPointCloud, PLYPointCloud, EaglePointCloud
A mesh	BunnyMesh, ArmadilloMesh, KnotMesh
Two clouds to register	DemoICPPointClouds, DemoColoredICPPointClouds, DemoFeatureMatchingPointClouds
An RGB-D sequence	SampleRedwoodRGBDImages, SampleFountainRGBDImages, BedroomRGBDImages, LoungeRGBDImages
A single RGB-D frame	SampleNYURGBDImage, SampleSUNRGBDImage, SampleTUMRGBDImage
A textured model	FlightHelmetModel, DamagedHelmetModel, MonkeyModel, AvocadoModel, CrateModel, SwordModel
A PBR texture set	WoodTexture, TilesTexture, MetalTexture, PaintedPlasterTexture, TerrazzoTexture, WoodFloorTexture
A full indoor scene	LivingRoomPointClouds, OfficePointClouds, RedwoodIndoorLivingRoom1/2, RedwoodIndoorOffice1/2

Full list: python -c "import open3d as o3d; print(dir(o3d.data))"