vis4d.op.detect.dense_anchor

Dense anchor-based head.

Functions

get_targets_per_batch(featmap_sizes, ...[, ...])	Get targets for all batch elements, all scales.
get_targets_per_image(target_boxes, anchors, ...)	Get targets per batch element, all scales.
images_to_levels(targets)	Convert targets by image to targets by feature level.

Classes

DenseAnchorHeadLoss(anchor_generator, ...[, ...])	Loss of dense anchor heads.
DenseAnchorHeadLosses(loss_cls, loss_bbox)	Dense anchor head loss container.
DetectorTargets(labels, label_weights, ...)	Targets for first-stage detection.

class DetectorTargets(labels: Tensor, label_weights: Tensor, bbox_targets: Tensor, bbox_weights: Tensor)

Targets for first-stage detection.

labels: Tensor

Alias for field number 0

label_weights: Tensor

Alias for field number 1

bbox_targets: Tensor

Alias for field number 2

bbox_weights: Tensor

Alias for field number 3

images_to_levels(targets)

Convert targets by image to targets by feature level.

Return type:

list[list[Tensor]]

get_targets_per_image(target_boxes, anchors, matcher, sampler, box_encoder, image_hw, target_class=1.0, allowed_border=0)

Get targets per batch element, all scales.

Parameters:

• target_boxes (Tensor) – (N, 4) Tensor of target boxes for a single image.

• anchors (Tensor) – (M, 4) box priors

• matcher (Matcher) – box matcher matching anchors to targets.

• sampler (Sampler) – box sampler sub-sampling matches.

• box_encoder (DeltaXYWHBBoxEncoder) – Encodes boxes into target regression parameters.

• image_hw (tuple[int, int]) – input image height and width.

• target_class (Tensor | float, optional) – class label(s) of target boxes. Defaults to 1.0.

• allowed_border (int, optional) – Allowed border for sub-sampling anchors that lie inside the input image. Defaults to 0.

Returns:

Targets, sum of positives, sum

of negatives.

Return type:

tuple[DetectorTargets, Tensor, Tensor]

get_targets_per_batch(featmap_sizes, target_boxes, target_class_ids, images_hw, anchor_generator, box_encoder, box_matcher, box_sampler, allowed_border=0)

Get targets for all batch elements, all scales.

Return type:

tuple[list[list[Tensor]], int]

class DenseAnchorHeadLosses(loss_cls: Tensor, loss_bbox: Tensor)

Dense anchor head loss container.

loss_cls: Tensor

Alias for field number 0

loss_bbox: Tensor

Alias for field number 1

class DenseAnchorHeadLoss(anchor_generator, box_encoder, box_matcher, box_sampler, loss_cls, loss_bbox, allowed_border=0)

Loss of dense anchor heads.

For a given set of multi-scale dense outputs, compute the desired target outputs and apply classification and regression losses. The targets are computed with the given target bounding boxes, the anchor grid defined by the anchor generator and the given box encoder.

__init__(anchor_generator, box_encoder, box_matcher, box_sampler, loss_cls, loss_bbox, allowed_border=0)

Creates an instance of the class.

Parameters:

• anchor_generator (AnchorGenerator) – Generates anchor grid priors.

• box_encoder (DeltaXYWHBBoxEncoder) – Encodes bounding boxes to the desired network output.

• box_matcher (Matcher) – Box matcher.

• box_sampler (Sampler) – Box sampler.

• loss_cls (TorchLossFunc) – Classification loss.

• loss_bbox (TorchLossFunc) – Bounding box regression loss.

• allowed_border (int) – The border to allow the valid anchor. Defaults to 0.

forward(cls_outs, reg_outs, target_boxes, images_hw, target_class_ids=None)

Compute RetinaNet classification and regression losses.

Parameters:

• cls_outs (list[Tensor]) – Network classification outputs at all scales.

• reg_outs (list[Tensor]) – Network regression outputs at all scales.

• target_boxes (list[Tensor]) – Target bounding boxes.

• images_hw (list[tuple[int, int]]) – Image dimensions without padding.

• target_class_ids (list[Tensor] | None, optional) – Target class labels.

Returns:

Classification and regression losses.

Return type:

DenseAnchorHeadLosses

__call__(cls_outs, reg_outs, target_boxes, images_hw, target_class_ids=None)

Type definition.

Return type:

DenseAnchorHeadLosses