12. Geometry Types¶

12.1 Point Clouds¶

Structure: unstructured vertices (no links).
Required arrays: vertices/. Optional companions: vertex_fragments/ (mandatory for multi-fragment chunks), vertex_attributes/<name>/, object_index/data (when objects exist, e.g. nuclei centroids), groups/data (region partitions).
links_convention: typically "implicit_sequential" — point clouds have no edges; the links/ array group is absent.
Spatial Indexing: direct spatial chunking; per-chunk fragmentation lets readers slice to single bins without decoding the whole chunk (see §6.3).
Multi-Resolution: per-object coarsening reduces each surviving object’s vertices to a single coarse-bin centroid metavertex per level.

Types: triangular (link_width = 3), quad, tetrahedral (link_width = 4), etc.
Storage options:
- Option 1 — Draco-encoded: vertices/<chunk>.encoding = "draco"; positions and faces co-encoded inside the Draco blob. links/0/ is omitted.
- Option 2 — Separate positions + faces: vertices/<chunk> raw float positions; links/0/<chunk> carries face records (link_width = 3 for triangles); link_fragments/<chunk> partitions faces by vertex fragment.
Cross-chunk faces: faces with three vertices in distinct chunks live in cross_chunk_links/0/data with link_width = 3 (replacing the dropped cross_chunk_faces/ array from pre-0.5).
Multi-Resolution: edge-collapse decimation or per-object metavertex aggregation; v0.7 chunk-scale growth keeps coarse-level chunk counts tractable.

Structure: graph of vertices and parent edges (tree).
Required arrays: vertices/, vertex_fragments/. Most skeletons additionally carry links/0/ (parent edges, link_width = 1).
links_convention:
- "explicit" — every parent edge materialised in links/0/<chunk>.
- "implicit_sequential_with_branches" (recommended) — sequential parents (parent_of(i) = i - 1) implicit; the links/0/<chunk> array stores only branch links and any links that cross fragment boundaries. Dramatically reduces storage (e.g. a 10k-vertex neuron with 50 branches stores ~50 link rows).
Per-vertex attributes: radius, vertex_type (soma / axon / dendrite), and so on — one row per vertex.
Cross-chunk parents: parent in chunk A, child in chunk B → one record in cross_chunk_links/0/data with endpoints ((A, parent_vi), (B, child_vi)).
Multi-Resolution: path simplification (Douglas-Peucker) keeps branch points and reduces straight-segment density.

Structure: ordered sequence of connected points.
Required arrays: vertices/, vertex_fragments/ (one fragment per streamline segment within each chunk). No links/ array.
links_convention: "implicit_sequential" — within a fragment, vertex i connects to i + 1.
Cross-chunk continuation: cross_chunk_links/0/data with link_width = 2 records connect a segment’s last vertex in chunk A to the next segment’s first vertex in chunk B.
Object identity: one object per full streamline; its manifest chains segments across chunks via mode-2 (explicit) manifest blocks if segments are shared across streamlines (shared_fragments), or mode-0 / mode-1 otherwise.
Multi-Resolution: point reduction along paths; pyramid levels carry fewer vertices per streamline.

Extensibility: writers may add a non-canonical entry to geometry_types — readers should treat unknown types as raw vertex + link arrays and skip conformance checks for that type.
Custom links: arbitrary link_width and arbitrary record semantics are supported by the underlying record format.
Validation: level-4 conformance checks (§13.1) apply only to the canonical geometry-type set; custom types bypass them.