tfm.vision.layers.MaskSampler

Samples and creates mask training targets.

tfm.vision.layers.MaskSampler(
    mask_target_size: int, num_sampled_masks: int, **kwargs
)

Methods

call

View source

call(
    candidate_rois: tf.Tensor,
    candidate_gt_boxes: tf.Tensor,
    candidate_gt_classes: tf.Tensor,
    candidate_gt_indices: tf.Tensor,
    gt_masks: tf.Tensor
)

Samples and creates mask targets for training.

Args
candidate_rois A tf.Tensor of shape of [batch_size, N, 4], where N is the number of candidate RoIs to be considered for mask sampling. It includes both positive and negative RoIs. The num_mask_samples_per_image positive RoIs will be sampled to create mask training targets.
candidate_gt_boxes A tf.Tensor of shape of [batch_size, N, 4], storing the corresponding groundtruth boxes to the candidate_rois.
candidate_gt_classes A tf.Tensor of shape of [batch_size, N], storing the corresponding groundtruth classes to the candidate_rois. 0 in the tensor corresponds to the background class, i.e. negative RoIs.
candidate_gt_indices A tf.Tensor of shape [batch_size, N], storing the corresponding groundtruth instance indices to the candidate_gt_boxes, i.e. gt_boxes[candidate_gt_indices[:, i]] = candidate_gt_boxes[:, i], where gt_boxes which is of shape [batch_size, MAX_INSTANCES, 4], M >= N, is the superset of candidate_gt_boxes.
gt_masks A tf.Tensor of [batch_size, MAX_INSTANCES, mask_height, mask_width] containing all the groundtruth masks which sample masks are drawn from. after sampling. The output masks are resized w.r.t the sampled RoIs.

Returns
foreground_rois A tf.Tensor of shape of [batch_size, K, 4] storing the RoI that corresponds to the sampled foreground masks, where K = num_mask_samples_per_image.
foreground_classes A tf.Tensor of shape of [batch_size, K] storing the classes corresponding to the sampled foreground masks.
cropoped_foreground_masks A tf.Tensor of shape of [batch_size, K, mask_target_size, mask_target_size] storing the cropped foreground masks used for training.