PointTransformerV3Encoder<T>
Models & Types · AiDotNet.ComputerVision.Segmentation.PointCloud
The Point Transformer V3 backbone (Wu et al., CVPR 2024), shared by PointTransformerV3
and Sonata: serialization, a sparse-convolution or linear stem, staged blocks of
serialized patch attention with xCPE, grid pooling between stages and, unless it is encoder-only,
a U-Net decoder of unpooling plus blocks.
Inheritance: LayerBase<T> → PointTransformerV3Encoder<T>
Implements: IShapeContract
For Beginners
This layer turns a point cloud into a feature vector for every point. It sorts the points along space-filling curves so that nearby points sit next to each other, lets each small group of neighbours attend to one another, and repeatedly merges neighbouring points into coarser clouds to see larger regions.
How It Works
Input: [points, width] for one cloud or [..., points, width] for a batch of
equally sized clouds. The first three channels are the xyz coordinates in metres; the features
are the remaining channels, or every channel when coordinates are also features (Sonata feeds
coordinate, colour and normal). Grid coordinates are floor((xyz - min) / gridSize) per
cloud, as Pointcept’s GridSample produces; the input is expected to be grid-sampled
already (one point per voxel), as in every PTv3 and Sonata configuration.
Output: one feature vector per input point, [..., points, OutputChannels]: the
decoder’s finest stage, or, encoder-only, Sonata’s feature up-casting of every stage concatenated
back to full resolution.
Only the features flow through T; grid coordinates, codes, orders, neighbour tables and
cluster maps are Int32/Int64 and coordinates are Double,
so a half-precision network never rounds an index.
Constructors
| Constructor | Summary |
|---|---|
PointTransformerV3Encoder(int inputChannels, bool coordinatesAsFeatures, double gridSize, int[] strides, int[] encoderDepths, int[] encoderChannels, int[] encoderNumHeads, int[] encoderPatchSizes, int[] decoderDepths, int[] decoderChannels, int[] decoderNumHeads, int[] decoderPatchSizes, double mlpRatio, double dropPathRate, bool enableConditionalPositionalEncoding, bool preNorm, bool shuffleOrders, SpaceFillingCurveSet serializationCurves, PointTransformerV3Revision revision, bool encoderOnly, bool maskToken, int? seed) |
Creates a PTv3 backbone. |
Properties
| Property | Summary |
|---|---|
bool IsEncoderOnly { get; } |
Whether this backbone has no decoder (Sonata). |
bool SupportsTraining { get; } |
|
int InputChannels { get; } |
Feature channels per point the layer expects after the coordinates. |
int InputWidthChannels { get; } |
Width of the input’s last axis: three coordinates plus the features, or just the features when they include the coordinates. |
int OutputChannels { get; } |
Feature channels per output point. |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
void ResetState() |