"""Computes the box loss (Fast-RCNN branch) of Mask-RCNN.
This function implements the box regression loss of the Fast-RCNN. As the
`box_outputs` produces `num_classes` boxes for each RoI, the reference model
expands `box_targets` to match the shape of `box_outputs` and selects only
the target that the RoI has a maximum overlap. (Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/roi_data/fast_rcnn.py) # pylint: disable=line-too-long
Instead, this function selects the `box_outputs` by the `class_targets` so
that it doesn't expand `box_targets`.
The box loss is smooth L1-loss on only positive samples of RoIs.
This function implements the mask loss of Mask-RCNN. As the `mask_outputs`
produces `num_classes` masks for each RoI, the reference model expands
`mask_targets` to match the shape of `mask_outputs` and selects only the
target that the RoI has a maximum overlap. (Reference: https://github.com/facebookresearch/Detectron/blob/master/detectron/roi_data/mask_rcnn.py) # pylint: disable=line-too-long
Instead, this implementation selects the `mask_outputs` by the `class_targets`
so that it doesn't expand `mask_targets`. Note that the selection logic is
done in the post-processing of mask_rcnn_fn in mask_rcnn_architecture.py.
Args:
mask_outputs: a float tensor representing the prediction for each mask,
with a shape of
[batch_size, num_masks, mask_height, mask_width].
mask_targets: a float tensor representing the binary mask of ground truth
labels for each mask with a shape of
[batch_size, num_masks, mask_height, mask_width].
select_class_targets: a tensor with a shape of [batch_size, num_masks],
representing the foreground mask targets.
Returns:
mask_loss: a float tensor representing total mask loss.
