PointTransformerV3Options

Options & Configuration · AiDotNet.ComputerVision.Segmentation.PointCloud

Configuration options for Point Transformer V3 semantic segmentation.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → PointTransformerV3EncoderOptions → PointTransformerV3Options

For Beginners

Leave these alone to build the network from the paper; change the arrays to make it smaller or larger.

How It Works

The defaults are Pointcept’s PT-v3m1 base configuration from the PTv3 paper (Wu et al., CVPR 2024, Table 9 / semseg-pt-v3m1-0-base): five encoder stages of depths (2, 2, 2, 6, 2) and widths (32, 64, 128, 256, 512), a four-stage decoder of depths (2, 2, 2, 2) and widths (64, 64, 128, 256), patch 1024, the four z/Hilbert orders, xCPE, pre-norm, MLP ratio 4 and drop path 0.3 – 46.2M backbone parameters.

Constructors

Constructor Summary
PointTransformerV3Options() Initializes a new instance with the PTv3 m1 base values.
PointTransformerV3Options(PointTransformerV3Options other) Initializes a new instance by copying from another instance.

Methods

Method Summary
void Validate() Throws when a value on this instance cannot produce a working model.