Neuroglancer precomputed (no extension)

The Neuroglancer precomputed mesh representation for segmented object surfaces: a small binary format (a vertex count, the vertex coordinates, then the triangle indices).

Format name	neuroglancer
Extensions	(none — pass file_format="neuroglancer")
Read / Write	✓ / ✓
Extra dependencies	—

Reading & writing

import meshioplusplus

mesh = meshioplusplus.read("mesh", file_format="neuroglancer")
meshioplusplus.neuroglancer.write("out", mesh)

write takes no keyword arguments. register_format is called with an empty extension list (register_format("neuroglancer", [], read, {"neuroglancer": write})), so this format is never auto-detected from a filename — file_format="neuroglancer" must be passed explicitly on both read and write.

File structure

Raw little-endian binary, no header/magic beyond the leading count:

1. <I (uint32) — number of vertices, n.
2. n * 3 × <f (float32) — flattened x0 y0 z0 x1 y1 z1 ... vertex coordinates.
3. The remainder of the file — flattened <I (uint32) triangle vertex indices, 3 per triangle; the remaining byte length must be a multiple of 12.

Read uses np.frombuffer rather than np.fromfile, specifically so the same code works transparently when handed a gzip.GzipFile-like file object, not just a plain path.

Cell types

triangle only.

Data mapping

None — no point_data, cell_data, or field_data; a bare Mesh(points, [CellBlock("triangle", ...)]).

Quirks & limitations

• Write silently casts points to float32 regardless of the input mesh's dtype — unlike most other meshio++ writers, there is no warning for this precision loss.
• Triangle-index validation rejects an index equal to or above the vertex count, in both engines.
• The C++ core reads and writes the format, so it is reachable from C, Fortran, Julia, R, WASM and the native CLI, always with an explicit format name; the two engines write the same bytes.

Notes

• tests/python/meshes/neuroglancer/simple1 (100 bytes, no file extension) — checked for ref_sum=20 and ref_num_cells=4.
• Implemented in both engines: src/cpp/src/formats/neuroglancer.cpp and neuroglancer/_neuroglancer.py; the registered shim tries the core first and falls back to the twin.