RandLA-Net

RandLA-Net classification and segmentation models.

First page of RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds

1911.11236 · November 2019

Classes:

RandLANetIntermediate –

Per-stage encoder features, positions and batch index, kept for the decoder skip connections.
LocalSpatialEncoding –

Per-edge spatial encoding MLP (RandLA-Net Section 3.2).
AttentivePooling –

Attention-weighted aggregation of pre-gathered edge features (RandLA-Net Section 3.3).
LocalFeatureAggregation –

Local Feature Aggregation module (RandLA-Net Section 3.4).
DilatedResidualBlock –

RandLA-Net dilated residual block (Fig. 3 of the paper).
RandLANetEncoder –

Stack of DilatedResidualBlock units interleaved with random-sampling
RandLANetDecoder –

Stack of PointNet2FeaturePropagation blocks specialized to RandLA-Net.
RandLANetClassification –

RandLA-Net classification model from
RandLANetSegmentation –

RandLA-Net segmentation model from

Functions:

random_max_pool –

Random sampling followed by a max-pool over the neighbors of each kept point.

RandLANetIntermediate ¶

Bases: NamedTuple

Per-stage encoder features, positions and batch index, kept for the decoder skip connections.

LocalSpatialEncoding ¶

LocalSpatialEncoding(
    in_channels: int,
    out_channels: int,
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

Per-edge spatial encoding MLP (RandLA-Net Section 3.2).

Wraps a single Linear+norm+act block that lifts an input feature to out_channels. Used twice per LocalFeatureAggregation: first on the raw 10-channel relative positional encoding, then on its output.

AttentivePooling ¶

AttentivePooling(
    in_channels: int,
    out_channels: int,
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

Attention-weighted aggregation of pre-gathered edge features (RandLA-Net Section 3.3).

Given \((E, C)\) edge features and a per-edge destination index, learn per-edge attention scores via a no-bias linear layer, softmax-normalize them across the neighbors of each destination point, sum the score-weighted features per destination, then project the result with a Linear+norm+act block.

Parameters:

in_channels (int) –

Channels of each edge feature.
out_channels (int) –

Output channels after the post-aggregation MLP.

LocalFeatureAggregation ¶

LocalFeatureAggregation(
    d_out: int,
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

Local Feature Aggregation module (RandLA-Net Section 3.4).

Stacks two LocalSpatialEncoding + AttentivePooling units to progressively grow the receptive field, doubling a per-point feature of dim \(d_\text{out} / 2\) to \(d_\text{out}\). Mirrors the LocSE + Attentive Pooling "dilated" combination in Fig. 3 of the paper. The 10-channel relative positional encoding follows the original QingyongHu/RandLA-Net channel order (cat([rel_dist, rel_xyz, xyz_i, xyz_j], dim=-1)) so pretrained weights load without permuting the first 1x1 kernel. The second LSE re-projects the output of the first LSE to match the original building_block.

DilatedResidualBlock ¶

DilatedResidualBlock(
    d_in: int,
    d_out: int,
    num_neighbors: int,
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

RandLA-Net dilated residual block (Fig. 3 of the paper).

Maps d_in channels to \(2 \cdot d_\text{out}\) via a residual path of MLP -> LocalFeatureAggregation -> MLP plus a parallel Linear+norm shortcut. The sum is activated by the configured activation. mlp2 and shortcut skip the activation (paper-mandated: Conv2d(activation=False) upstream); the configured activation is applied once after the residual sum.

Parameters:

d_in (int) –

Number of input channels.
d_out (int) –

"Configuration" channel count; the block actually outputs \(2 \cdot d_\text{out}\).
num_neighbors (int) –

Number of neighbors for the local feature aggregation.

RandLANetEncoder ¶

RandLANetEncoder(
    in_channels: int,
    encoder_channels: Sequence[int],
    decimation: Union[int, Sequence[int]] = 4,
    num_neighbors: Union[int, Sequence[int]] = 16,
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

Stack of DilatedResidualBlock units interleaved with random-sampling K-NN max-pool decimation (RandLA-Net Section 3 / Fig. 2).

Each encoder block doubles its \(d_\text{out}^\text{config}\) to produce a \(2 \cdot d_\text{out}^\text{config}\) channel feature; the per-stage decimation ratio then sub-samples the cloud by that factor.

When return_intermediates=True is passed to forward, the encoder returns the bottleneck features plus the per-stage skip features in fine-to-coarse order: intermediates[0] is block 0's PRE-decimation output (used as the full-resolution skip in RandLANetDecoder); subsequent entries are the POST-decimation outputs of blocks \(0 \ldots N-2\). The decoder consumes them in reverse.

Parameters:

in_channels (int) –

Number of input channels.
encoder_channels (Sequence[int]) –

Output channels of each dilated residual block (must be even).
decimation (Union[int, Sequence[int]], default: 4 ) –

Decimation factor between consecutive encoder blocks. Either a single int or a per-block sequence of length \(N\).
num_neighbors (Union[int, Sequence[int]], default: 16 ) –

Number of neighbors for the K-NN graph in each block. Either a single int or a per-block sequence of length \(N\).

RandLANetDecoder ¶

RandLANetDecoder(
    in_channels: int,
    skip_channels: Sequence[int],
    fp_channels: Sequence[int],
    *,
    act: Union[str, Callable, None],
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None],
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: Module

Stack of PointNet2FeaturePropagation blocks specialized to RandLA-Net.

Each block does 1-NN upsampling from the deeper resolution to its pos_skip resolution, concatenates with the encoder skip features, then applies a single linear+norm+act projection. The first block consumes the bottleneck features and the deepest encoder skip; subsequent blocks each cut the resolution by a stage of decimation until the finest (full-resolution) skip is reached.

Note

Upstream RandLA-Net cats [skip, interp] while PointNet2FeaturePropagation cats [interp, skip]. To preserve weight equivalence with the upstream checkpoint, the conversion utilities (convert_randlanet_state_dict, convert_open3d_randlanet_state_dict) swap the first linear layer's column blocks per FP block.

Parameters:

in_channels (int) –

Channels at the bottleneck (input to the first FP).
skip_channels (Sequence[int]) –

Per-stage encoder skip channels in coarse-to-fine order.
fp_channels (Sequence[int]) –

Per-stage decoder output channels (same length as skip_channels).
act (Union[str, Callable, None]) –

Activation passed to each PointNet2FeaturePropagation MLP.
act_kwargs (Optional[Dict[str, Any]], default: None ) –

Activation kwargs.
act_first (bool, default: False ) –

If True, activation is applied before normalization.
norm (Union[str, Callable, None]) –

Normalization passed to each PointNet2FeaturePropagation MLP.
norm_kwargs (Optional[Dict[str, Any]], default: None ) –

Normalization kwargs.
bias (bool, default: False ) –

Whether to use bias in the MLP layers.

Note

No paper-specific defaults are baked in here; callers (typically the registered model factory) supply them.

RandLANetClassification ¶

RandLANetClassification(
    in_channels: int,
    num_classes: int,
    *,
    stem_channels: Optional[int] = 8,
    encoder_channels: Sequence[int],
    decimation: Union[int, Sequence[int]] = 4,
    num_neighbors: Union[int, Sequence[int]] = 16,
    aggr_channels: Optional[
        Union[int, Sequence[int]]
    ] = None,
    dropout: float = 0.0,
    global_pool: PoolLike = "max",
    act: Union[str, Callable, None] = "relu",
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None] = "batch_norm",
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: ClassificationModel

RandLA-Net classification model from RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds by Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, Andrew Markham.

Random sampling for downsampling and dilated residual blocks with local feature aggregation; point features are pooled globally after the encoder for classification.

Parameters:

in_channels (int) –

Number of input channels.
num_classes (int) –

Number of classes.
stem_channels (Optional[int], default: 8 ) –

Number of channels in the stem MLP. Set to None to skip the stem.
encoder_channels (Sequence[int]) –

Output channels of each dilated residual block (must be even).
decimation (Union[int, Sequence[int]], default: 4 ) –

Decimation factor between consecutive encoder blocks.
num_neighbors (Union[int, Sequence[int]], default: 16 ) –

Number of neighbors for the kNN graph in each block.
aggr_channels (Optional[Union[int, Sequence[int]]], default: None ) –

Optional channels for the aggregation MLP applied before global pooling.
dropout (float, default: 0.0 ) –

Dropout rate before the classification head.
global_pool (PoolLike, default: 'max' ) –

Global pooling operation.

Methods:

configure_stem –

Build the stem lifting the input features to stem_channels, or None when stem_channels is unset.
configure_encoder –

Build the RandLANetEncoder backbone.
configure_aggr –

Build the aggregation MLP applied to the encoder output, or None when aggr_channels is unset.

Attributes:

num_features (int) –

Feature dimension \(C\) of the encoder output, after the optional aggregation MLP.

num_features `property` ¶

num_features: int

Feature dimension \(C\) of the encoder output, after the optional aggregation MLP.

configure_stem ¶

configure_stem() -> Optional[MLP]

Build the stem lifting the input features to stem_channels, or None when stem_channels is unset.

configure_encoder ¶

configure_encoder() -> RandLANetEncoder

Build the RandLANetEncoder backbone.

configure_aggr ¶

configure_aggr() -> Optional[MLP]

Build the aggregation MLP applied to the encoder output, or None when aggr_channels is unset.

RandLANetSegmentation ¶

RandLANetSegmentation(
    in_channels: int,
    num_classes: int,
    *,
    stem_channels: Optional[int] = 8,
    encoder_channels: Sequence[int],
    fp_channels: Sequence[int],
    head_channels: Sequence[int] = (64, 32),
    decimation: Union[int, Sequence[int]] = 4,
    num_neighbors: Union[int, Sequence[int]] = 16,
    aggr_channels: Optional[
        Union[int, Sequence[int]]
    ] = None,
    dropout: float = 0.5,
    act: Union[str, Callable, None] = "relu",
    act_kwargs: Optional[Dict[str, Any]] = None,
    act_first: bool = False,
    norm: Union[str, Callable, None] = "batch_norm",
    norm_kwargs: Optional[Dict[str, Any]] = None,
    bias: bool = False,
)

Bases: SegmentationModel

RandLA-Net segmentation model from RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds by Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, Andrew Markham.

Encoder uses random sampling between dilated residual blocks; decoder uses 1-NN nearest-neighbor interpolation with concatenation skips and a single 1x1 linear+BN+act per stage. The skip used at full resolution is the pre-decimation output of the first encoder block.

Parameters:

in_channels (int) –

Number of input channels.
num_classes (int) –

Number of classes.
stem_channels (Optional[int], default: 8 ) –

Number of channels in the stem MLP. Set to None to skip the stem.
encoder_channels (Sequence[int]) –

Output channels of each dilated residual block (must be even).
fp_channels (Sequence[int]) –

Per-stage decoder channels (one list per upsampling step).
head_channels (Sequence[int], default: (64, 32) ) –

Hidden channels of the segmentation head MLP.
decimation (Union[int, Sequence[int]], default: 4 ) –

Decimation factor between consecutive encoder blocks.
num_neighbors (Union[int, Sequence[int]], default: 16 ) –

Number of neighbors for the kNN graph in each block.
aggr_channels (Optional[Union[int, Sequence[int]]], default: None ) –

Channels for the bottleneck MLP between the encoder and the decoder (the upstream "decoder_0").
dropout (float, default: 0.5 ) –

Dropout rate inside the segmentation head MLP.
act (Union[str, Callable, None], default: 'relu' ) –

Activation type for both the decoder FP MLPs and the segmentation head MLP (string passed to create_act, or a Callable / nn.Module).
act_kwargs (Optional[Dict[str, Any]], default: None ) –

Keyword arguments forwarded to the activation.
act_first (bool, default: False ) –

If True, apply activation before normalization (PyG's MLP act_first semantics): Linear → Act → Norm → Dropout instead of the default Linear → Norm → Act → Dropout.
norm (Union[str, Callable, None], default: 'batch_norm' ) –

Normalization type for the decoder FP MLPs and the head MLP.
norm_kwargs (Optional[Dict[str, Any]], default: None ) –

Keyword arguments forwarded to the normalization layers.
bias (bool, default: False ) –

Whether the decoder/head hidden linear layers carry an explicit bias. The final head layer always uses bias=True since it has no normalization.

Methods:

configure_stem –

Build the stem lifting the input features to stem_channels, or None when stem_channels is unset.
configure_encoder –

Build the RandLANetEncoder backbone.
configure_aggr –

Build the aggregation MLP applied to the encoder output, or None when aggr_channels is unset.
configure_decoder –

Build the RandLANetDecoder upsampling the bottleneck features back through the encoder skips.

Attributes:

num_features (int) –

Feature dimension \(C\) of the decoder output.

num_features `property` ¶

num_features: int

Feature dimension \(C\) of the decoder output.

configure_stem ¶

configure_stem() -> Optional[MLP]

Build the stem lifting the input features to stem_channels, or None when stem_channels is unset.

configure_encoder ¶

configure_encoder() -> RandLANetEncoder

Build the RandLANetEncoder backbone.

configure_aggr ¶

configure_aggr() -> Optional[MLP]

Build the aggregation MLP applied to the encoder output, or None when aggr_channels is unset.

configure_decoder ¶

configure_decoder() -> RandLANetDecoder

Build the RandLANetDecoder upsampling the bottleneck features back through the encoder skips.

random_max_pool ¶

random_max_pool(
    x: Tensor,
    pos: Tensor,
    batch: Tensor,
    factor: int,
    num_neighbors: int,
    generator: Optional[Generator] = None,
) -> Tuple[Tensor, Tensor, Tensor]

Random sampling followed by a max-pool over the neighbors of each kept point.

Parameters:

x (Tensor) –

Packed point features, shape \((N, C)\).
pos (Tensor) –

Packed point positions, shape \((N, 3)\).
batch (Tensor) –

Per-point batch index, shape \((N,)\).
factor (int) –

Decimation factor, keeping \(N // \text{factor}\) points per cloud.
num_neighbors (int) –

Number of neighbors gathered around each kept point.
generator (Optional[Generator], default: None ) –

Generator driving the random sampling.

Returns:

Tuple[Tensor, Tensor, Tensor] –

The pooled features, positions and batch index of the kept points.

RandLA-Net

RandLANetIntermediate ¶

LocalSpatialEncoding ¶

AttentivePooling ¶

LocalFeatureAggregation ¶

DilatedResidualBlock ¶

RandLANetEncoder ¶

RandLANetDecoder ¶

RandLANetClassification ¶

num_features property ¶

configure_stem ¶

configure_encoder ¶

configure_aggr ¶

RandLANetSegmentation ¶

num_features property ¶

configure_stem ¶

configure_encoder ¶

configure_aggr ¶

configure_decoder ¶

random_max_pool ¶

num_features `property` ¶

num_features `property` ¶