lingvo.tasks.car.pillars module

PointPillars implementation.

[1] PointPillars. https://arxiv.org/abs/1812.05784

lingvo.tasks.car.pillars.SparseToDense(grid_shape, locations, feats)

Converts a sparse representation back to the dense grid.

Parameters

• grid_shape – (nx, ny, nz). The shape of the grid.

• locations – [b, p, 3]. Locations of the pillars.

• feats – [b, p, fdims]. Extracted features for pillars.

Returns

grid_feats of shape [b, nx, ny, nz * fdims].

class lingvo.tasks.car.pillars.PointsToGridFeaturizer(*args, **kwargs)

Bases: lingvo.core.base_layer.BaseLayer

Layer for processing points to grid outputs.

classmethod Params(num_laser_features, num_output_features=64)

Returns the layer params.

FProp(theta, input_batch)

Compute features for the pillars and convert them back to a dense grid.

Parameters

• theta – A NestedMap object containing variable values of this task.

• input_batch –

  A NestedMap object containing input tensors. Following keys are required:

  • grid_num_points: Integer tensor with shape [batch size, nx, ny, nz], where nx, ny, nz corresponds to the grid sizes (i.e., number of voxels in each axis dimension).

  • pillar_points: Float tensor with shape [batch size, num_pillars, num_points_per_pillar, 3 + num_laser_features]

  • pillar_centers: Float tensor with shape [batch size, num_pillars, num_points_per_pillar, 3]

  • pillar_locations: Float tensor with shape [batch size, num_pillars, 3]

Returns

The dense features with shape [b, nx, ny, nz * fdims].

class lingvo.tasks.car.pillars.Builder

Bases: lingvo.tasks.car.builder_lib.ModelBuilderBase

Builder for the Pillars model.

Featurizer(name, dims)

_Deconv(name, filter_shape, stride)

_Block(name, stride, repeats, idims, odims)

[1]. Sec 2.2.

_TopDown(name, strides=(2, 2, 2), channel_multiplier=1)

[1]. Sec 2.2.

_Upsample(name, stride, idims, odims)

[1]. Sec 2.2.

Contract(down_strides=(2, 2, 2), channel_multiplier=1)

Contracting part of [1] Sec 2.2.

Expand(odims, channel_multiplier=1)

Expanding part of [1] Sec 2.2.

Backbone(odims, down_strides=(2, 2, 2), channel_multiplier=1)

[1]. Sec 2.2.

Detector(name, idims, odims, conv_init_method=None, bias_params_init=None)

class lingvo.tasks.car.pillars.LossNormType(value)

Bases: enum.Enum

An enumeration.

NO_NORM = 0

NORM_BY_NUM_POSITIVES = 1

class lingvo.tasks.car.pillars.ModelV1(*args, **kwargs)

Bases: lingvo.tasks.car.point_detector.PointDetectorBase

PointPillars model.

Base class implements common Decoder functions, though they can be overridden if desired.

NUM_OUTPUT_CHANNELS = 128

classmethod Params(grid_size_z=1, num_anchors=2, num_classes=1, num_laser_features=1)

Returns the layer params.

ComputePredictions(theta, input_batch)

Computes predictions for input_batch.

Parameters

• theta – A NestedMap object containing variable values of this task.

• input_batch – A NestedMap object containing input tensors to this tower.

Returns

A NestedMap contains

logits - [b, nx, ny, nz, na, 7 + num_classes]

_ComputeClassificationLoss(predictions, input_batch, class_weights)

Compute classification loss for the given predictions.

Parameters

• predictions – The output of ComputePredictions, contains: logits - [b, nx, ny, nz, na, 7 + num_classes]. na is the number of anchor boxes per cell. […, :7] are (dx, dy, dz, dw, dl, dh, dt).

• input_batch – The input batch from which we accesses the groundtruth.

• class_weights – Per-class weights to use in loss computation.

Returns

Classification loss.

ComputeLoss(theta, predictions, input_batch)

Computes loss and other metrics for the given predictions.

Parameters

• theta – A NestedMap object containing variable values of this task.

• predictions – The output of ComputePredictions, contains: logits - [b, nx, ny, nz, na, 7 + num_classes]. na is the number of anchor boxes per cell. […, :7] are (dx, dy, dz, dw, dl, dh, dt).

• input_batch – The input batch from which we accesses the groundtruth.

Returns

Two dicts defined as BaseTask.ComputeLoss.

_BBoxesAndLogits(input_batch, predictions)

Decode an input batch, computing predicted bboxes from residuals.