Transforms¶

A transform takes one sample dict and returns a new one. Compose chains them (RandomApply chains them under a probability), and a dataset applies the chain to every sample it loads. The API follows MONAI's dict transforms and PyTorch Geometric's Data conventions.

Each transform reads and writes the keys you name, so a sample can carry any extra key (intensity, color, ...) without every transform handling it.

import torch
import torch_pointcloud.transforms as T

pos = torch.randn(2048, 3)
color = torch.rand(2048, 3)

pipeline = T.Compose([
    T.Rescale(keys="pos", method="centroid"),
    T.Shift(keys="pos", method="bbox", axes=[0, 1]),
    T.RandomSample(keys=("pos", "color"), num_samples=1024),
])

scene = pipeline({"pos": pos, "color": color})

Pretrained checkpoints

A pretrained checkpoint carries its own preprocessing in info["transform"], for inference on its dataset.

Sampling and downsampling¶

	Transform	Description
	`RandomSample`	Uniform random subsample with shared indices across keys
	`FarthestPointSample`	FPS subsample (well-distributed)
	`RandomSampleFaceVertices`	Sample points on a mesh's faces
	`Voxelize`	Voxel-grid downsample with per-voxel reduction

Sampling keys¶

A transform that changes the number of points keeps pos, x, segment and batch aligned at the new resolution, and records how to get back when given the dst_* keys: inverse from Voxelize and DivisiblePad (dst_inverse_key), index from the selection samplers (dst_index_key), next to the origin_pos copy the registered pipelines keep. The sampling keys of the data format page define the four keys and how they compose.

Geometry / shifting¶

	Transform	Description
	`Shift`	Subtract a computed offset (`bbox`, `centroid`, or `min`)
	`AxisMinOffset`	Per-point offset from axis minimum (height feature)

Scaling / normalization¶

	Transform	Description
	`Rescale`	Center and rescale to unit extent (`centroid` / `bbox` / `centroid_extent` / `min_sphere`)
	`Normalize`	Per-channel \((x - \mu) / \sigma\) standardization
	`MaxAbsScale`	Per-channel division by the largest absolute value of the sample

Masking and filtering¶

	Transform	Geometry
	`BoxMask`	Axis-aligned bounding box (AABB)
	`CubeMask`	L∞ / Chebyshev ball (hypercube)
	`SphereMask`	L2 / Euclidean ball
	`ApplyMask`	Apply any precomputed mask to one or more keys
	`RemoveNearOrigin`	One-shot L2 filter around origin

Utilities¶

	Transform	Description
	`Cat`	Concatenate multiple keys' tensors along a dim
	`CopyItems`	Clone a key's value under a new name
	`RenameItems`	Move a key to a new name
	`KeepItems`	Drop everything not in a whitelist
	`SetValue`	Set keys to literal values
	`Reduce`	Reduce a tensor along a dim (`min`/`max`/`mean`/`sum`)
	`OneHot`	One-hot encode integer labels
	`Relabel`	Remap integer labels via a lookup table
	`Scale`	Multiply by a scalar
	`Divide`	Divide by a scalar
	`DivideItems`	`data[k] / data[div_k]` element-wise
	`SubtractItems`	`data[k] - data[sub_k]` element-wise
	`Abs`	Element-wise absolute value
	`ToFloat`	Cast tensors to float32
	`ToTensor`	Convert lists / arrays to tensors
	`ToDevice`	Move to a device
	`OnesLike`	Add a key whose tensor is `torch.ones_like(...)`