PCD (.pcd)

The Point Cloud Library format (v0.7): a short text header (VERSION, FIELDS, SIZE, TYPE, COUNT, WIDTH, HEIGHT, VIEWPOINT, POINTS, DATA) followed by the points as ASCII rows (DATA ascii), packed little-endian records (DATA binary) or a field-by-field (struct-of-arrays) block inside an LZF stream (DATA binary_compressed, two uint32 sizes then the stream).

Format name	pcd
Extensions	.pcd
Read / Write	✓ / ✓
Extra dependencies	—

Reading & writing

import meshioplusplus

mesh = meshioplusplus.read("cloud.pcd")
meshioplusplus.pcd.write("out.pcd", mesh, data="binary")

• data — "ascii", "binary" (the default) or "binary_compressed"; when omitted, binary=True/False picks between the first two.
• point_dtype — the precision of x y z (and of normals, curvature and intensity): "float32" (the default: PCL's typed PointXYZ loaders reject any other field type), "float64", or "keep" to follow the mesh's own points (what the in-place ascii/binary/compress/decompress verbs use). A lossy narrowing is recorded in the provenance.
• drop_invalid (read) — drop the points whose x, y or z is not finite (an organised cloud's invalid returns) and, when any is dropped, the organisation with them.

Mesh mapping

The points plus exactly one vertex block — what subsample_points produces. Point data is keyed by field name, with three conventions: normal_x/y/z → "normals" (n, 3); PCL's rgb/rgba, a uint32 0x00RRGGBB/0xAARRGGBB that lives in a float32 slot (unpacked by bit-cast, never by value) → "rgb" (n, 3) / "rgba" (n, 4) uint8; any other field keeps its own name and dtype, with COUNT > 1 giving (n, count). _ padding fields are skipped. x/y/z keep the file's precision (all F4 → float32, otherwise float64).

File structure

# .PCD v0.7 - Point Cloud Data file format
VERSION 0.7
FIELDS x y z [...]
SIZE 4 4 4 [...]
TYPE F F F [...]
COUNT 1 1 1 [...]
WIDTH <n>
HEIGHT 1
VIEWPOINT <tx ty tz qw qx qy qz>
POINTS <n>
DATA ascii | binary | binary_compressed
<points>

Quirks & limitations

• Organised clouds (HEIGHT > 1) are kept whole, NaN rows included, with WIDTH/HEIGHT in field_data["pcd:width"]/["pcd:height"] (restored on write when WIDTH * HEIGHT equals the point count).
• VIEWPOINT is recorded, never applied: a non-identity viewpoint goes to field_data["pcd:viewpoint"].
• Only vertex cells are writable; anything else, all cell data and unrelated field data are dropped with a warning and a provenance note.
• Point-data names are sanitised to the header's word vocabulary (spaces become _, collisions get a numeric suffix).

Notes

• tests/python/meshes/pcd/ — PCL's own test corpus (BSD-3, see LICENSE.PCL there): bun0.pcd (ascii, normals + curvature), colored_cloud.pcd (binary, organised, uint rgb), pcl_logo.pcd (binary_compressed, float-slot rgb, non-identity viewpoint), milk_color.pcd (binary_compressed, rgba).
• The C++ core handles all three DATA modes. Its LZF codec is an independent implementation of the liblzf stream format (no liblzf code is copied), and the Python reference writes byte-identical files. The CLI verbs ascii, binary, compress (binary_compressed) and decompress rewrite a .pcd in place, and the MCP convert tool takes mode and compression: lzf. The flat bindings (C, Fortran, Julia, R, WASM) read every mode and write ascii/binary; binary_compressed writing is on the roadmap.

Web output

A PCD cloud is a vertex block (or no cells at all), which the glTF writer turns into a POINTS primitive, keeping the normals point data as NORMAL and every other point array as a raw _NAME attribute: meshioplusplus convert cloud.pcd cloud.glb puts a scanned cloud on the web. Run compute_normals first when the cloud has none.