Point cloud annotation¶

The Point Cloud Annotation plugin labels LiDAR points with ASPRS classes directly on the map. Use it to fix misclassified points (for example, stadium seating left as "never classified") or to build training labels for point cloud segmentation models. It is the first slice of the plan tracked in opengeos/GeoLibre#2749.

Start a session¶

Load a point cloud with Add Data → LiDAR Layer (LAS, LAZ, COPC, or EPT).
Zoom to the area you want to label. A COPC or EPT cloud streams by level of detail, so the session holds the points loaded at the current view. For a streamed COPC, Full detail in view loads every point inside the view at the file's full resolution (up to 4 million points) and keeps it loaded however you zoom, so labels go on every point; Release full detail lets the area stream normally again. It works before or during a session; during one, the area must fit in the space streaming left free. A large area can take a minute or two to download.
Open Plugins → Point Cloud Annotation, pick the cloud, and choose Start annotating.

While a session runs, the plugin colours the cloud by classification and pauses streaming, so no edited point is dropped when you move the map. Scroll to zoom and right-drag to tilt or rotate as usual.

Select and assign¶

Box (B), Lasso (L) and Brush (P): drag on the map to select points. The brush selects everything within its radius of the dragged path; set its size in the panel or with [ and ]. Hold Shift to add to the selection or Alt to remove from it, or pick a mode in the panel. Pan returns left-drag to moving the map.
Polygon (G): click vertices, then double-click, press Enter or click the first vertex to close the ring and select what it encloses. Esc cancels. Hold Shift or Alt on the closing click to add or subtract.
Selection goes through all depths, like a camera frustum. Narrow it with the Min Z / Max Z filter (metres) or Only points in class, which relabels just the points currently in one class.
Lock a class in Classes in session to protect its points: locked classes, hidden classes (toggled in the LiDAR panel's legend) and points outside the LiDAR panel's elevation filter are never selected.
Selected points are drawn in yellow. Choose the class to assign (or press 0-9 for ASPRS classes 0-9) and press Apply (Enter); Clear (Esc) drops the selection. Clicking a row in Classes in session makes it the class to assign.
Undo and Redo step through the class assignments of the session.

Custom classes¶

Open Custom classes under the assign controls to add classes of your own (for example Car or Solar panel). Give each a code from 19 to 255 (ASPRS reserves 19-63 for future standard classes and leaves 64-255 to users, so the panel suggests the first free code from 64), a name and a colour, then press Add. A custom class can be assigned, locked and boxed like a standard one; the LiDAR panel's legend and tooltips show its name and colour. Custom classes are saved with the project, and the exports write their codes (the LAS classification byte holds 0-255). A notebook or agent can define them ahead of time with Map.set_point_cloud_classes or the set_point_cloud_classes MCP tool.

Instances¶

Instances group points that belong to one object, such as one car or one tree, for instance segmentation. Select its points and press New instance (N): they get the class chosen under Assign class and a new instance id, as one undoable edit. Assign class to points on a 3D box does the same for the points inside the box. Instances lists each one with its most common class and point count; Select selects its points again and Dissolve removes the instance while its points keep their class. Instance ids are saved with the project like classes, and exported with the points (see below).

Polygons to points¶

Polygons to points lifts a polygon layer straight up through the cloud: masks from Plugins → SamGeo segmenting imagery, building footprints, or any other polygons in the project. Pick the layer, then:

Select inside selects the points whose position falls inside any polygon (holes excluded), combined with the current selection like a drawn shape.
Instance per polygon gives the points inside each polygon the class to assign and their own instance, as one undoable edit. A point in overlapping polygons belongs to the first one.

A footprint takes every point above and below it, ground included, so narrow it with the Min Z / Max Z and Only points in class filters first (for example, only unclassified points above the ground). Locked, hidden and elevation-filtered points are skipped, as in a drawn selection.

Pre-labelling with Whitebox¶

Pre-label (Whitebox) runs a Whitebox LiDAR classifier on the session's points in the browser (the same WebAssembly build as Processing → Whitebox) and applies its classes as one undoable edit:

Ground (improved ground point filter) separates ground (2) from everything else (1).
Ground, vegetation and unclassified (classify LiDAR) also marks vegetation.

Only relabel unclassified points (0 and 1), on by default, keeps every point you or the survey already classified. Locked and hidden classes are never changed. The tool runs in tiles of about 750,000 points with a 20 m overlap, since the WebAssembly build runs out of memory on a few million points at once; a 4.6 million point session takes a couple of minutes. The result is a starting point: review it, then correct it with the selection tools.

Notebooks and agents can pre-label without the app: Map.prelabel_point_cloud in the Python package and the prelabel_point_cloud MCP tool run the same classifiers on a local copy of the layer's file (with the geolibre[pointcloud] extra) and save the result as the layer's labels, which the annotator shows when the project opens.

3D boxes (cuboids)¶

Label objects such as buildings, trees or vehicles with oriented 3D boxes:

Auto box (A): click a point on an object. The annotator grows a cluster from it through points within 0.75 m of each other, skipping ground, water, noise and locked or hidden classes, and fits the tightest box around it: the minimum-area rectangle of the cluster's footprint, rotated to the object, and its full height.
Box from selection: fits a box to the points you selected with the box, lasso or brush tools.

New boxes take the class chosen under Assign class. Each box in Objects (3D boxes) has its own class and three actions: Select points (the points inside it), Assign class to points (undoable like Apply; the points also become a new instance) and Delete. Click a box to open Box views, three orthographic views of the points around it: top, side (along its length) and front (across it). Drag inside the box to move it, drag an edge to resize it (the opposite face stays put), drag the knob beyond the front edge in the top view to rotate it, and scroll to zoom about the pointer. While a box is open in the views, the keyboard nudges it: arrow keys move it 10 cm north/south/east/west (1 m with Shift), Q and E rotate it by 1° (5° with Shift), and + / - raise or lower its top by 10 cm.

Each box also has a review status (New, Reviewed or Flagged) and free-form Attributes: name/value pairs such as make: Ford or occluded: yes, up to 32 per box. Expand Attributes, type a name and value and press Add; edit a value in place, or Remove it.

Boxes on a cloud loaded from a URL are saved with the project, with their status and attributes. Export them as Boxes as GeoJSON (one footprint polygon per box, with its class, z_min, z_max, size, heading, status and attributes as properties) or Boxes as Segments.ai JSON (a pointcloud-cuboid label in the same CRS and units as the LAS export, with the heading measured from grid east, and the status and attributes in each annotation's attributes).

3D vectors¶

Trace linear and point features in 3D: kerbs, lane markings and power lines as Polylines (V), roof outlines and footprints as 3D polygons, and poles or tree tops as Keypoints (K). Click points on the cloud; each vertex snaps to the nearest drawn point within 12 pixels and takes its elevation, so the vector follows the surface. Double-click or press Enter to finish a polyline or polygon, Backspace removes the last vertex and Esc cancels. The map's zoom and rotation are held while a vector is being drawn. Vectors take the class chosen under Assign class; change it, or delete a vector, in the 3D vectors list.

Vectors on a cloud loaded from a URL are saved with the project. Export them as Vectors as GeoJSON (3D Point, LineString and Polygon features in WGS 84 with the class and 3D length) or Vectors as Segments.ai JSON (a pointcloud-vector label with polyline, polygon and point annotations in the same CRS and units as the LAS export).

Saving labels with the project¶

Labels on a cloud loaded from a URL are saved with the project. Each edit is stored against the point's source node and its index in that node, not its position in memory, so the labels are re-applied when the project reopens and as streamed nodes load again, whatever order they arrive in. Labels on a local file are not saved (the file cannot be reopened from the project), so export them.

Export¶

LAS 1.4 writes every point loaded in the session (point format 7 with RGB, or 6 without) with its edited class, intensity, returns, GPS time and scan angle. Coordinates are written back in the source file's CRS, feet included, when its WKT is known, and in WGS 84 otherwise. When the session has instances, each point also carries an instance extra-bytes dimension (uint32, 0 for none), which laspy, PDAL and LAStools read by name.
LAZ (compressed) writes the same records as the LAS export, compressed with LASzip in the browser (a laz-rs WebAssembly build). It is typically about a quarter of the LAS size.
NumPy (.npy) writes a structured array with x, y, z (float64, in the same CRS as the LAS export), intensity, classification, red, green, blue when the cloud has colour, and instance when the session has instances, ready for numpy.load.
Segments.ai JSON writes a pointcloud-segmentation label. Its point_annotations line up point for point with the LAS file, with one annotation per instance and one per class for points in no instance, each with category_id set to the class code.

The exports above cover the points loaded in the session. Full file with labels writes a whole local copy of the point cloud with every saved label (and instance id) applied instead: enter the path of the local copy (the same LAS, LAZ or COPC file the layer streams) and an output .las or .laz path, then Write file. It runs in the GeoLibre server, which reads the file chunk by chunk, so files larger than memory work; a COPC is written as plain LAS/LAZ. The section shows when the server is running and the cloud was loaded from a URL (only those have saved labels). In the web build the paths must be inside the server's allowed folders.

Finish session resumes streaming.

Limitations¶

MapLibre renderer only.
COPC output is not available; export LAZ and convert it with PDAL (writers.copc) if you need a COPC file.
Boxes rotate about the vertical only (no pitch or roll).
Instances are not colored by id on the map; select one from Instances to see its points highlighted.