tf.raw_ops.GenerateBoundingBoxProposals

This op produces Region of Interests from given bounding boxes(bbox_deltas) encoded wrt anchors according to eq.2 in arXiv:1506.01497

tf.raw_ops.GenerateBoundingBoxProposals(
    scores, bbox_deltas, image_info, anchors, nms_threshold, pre_nms_topn, min_size,
    post_nms_topn=300, name=None
)
  The op selects top `pre_nms_topn` scoring boxes, decodes them with respect to anchors,
  applies non-maximal suppression on overlapping boxes with higher than
  `nms_threshold` intersection-over-union (iou) value, discarding boxes where shorter
  side is less than `min_size`.
  Inputs:
  `scores`: A 4D tensor of shape [Batch, Height, Width, Num Anchors] containing the scores per anchor at given postion
  `bbox_deltas`: is a tensor of shape [Batch, Height, Width, 4 x Num Anchors] boxes encoded to each anchor
  `anchors`: A 1D tensor of shape [4 x Num Anchors], representing the anchors.
  Outputs:
  `rois`: output RoIs, a 3D tensor of shape [Batch, post_nms_topn, 4], padded by 0 if less than post_nms_topn candidates found.
  `roi_probabilities`: probability scores of each roi in 'rois', a 2D tensor of shape [Batch,post_nms_topn], padded with 0 if needed, sorted by scores.

Args:

  • scores: A Tensor of type float32. A 4-D float tensor of shape [num_images, height, width, num_achors] containing scores of the boxes for given anchors, can be unsorted.
  • bbox_deltas: A Tensor of type float32. A 4-D float tensor of shape [num_images, height, width, 4 x num_anchors]. encoding boxes with respec to each anchor. Coordinates are given in the form [dy, dx, dh, dw].
  • image_info: A Tensor of type float32. A 2-D float tensor of shape [num_images, 5] containing image information Height, Width, Scale.
  • anchors: A Tensor of type float32. A 2-D float tensor of shape [num_anchors, 4] describing the anchor boxes. Boxes are formatted in the form [y1, x1, y2, x2].
  • nms_threshold: A Tensor of type float32. A scalar float tensor for non-maximal-suppression threshold.
  • pre_nms_topn: A Tensor of type int32. A scalar int tensor for the number of top scoring boxes to be used as input.
  • min_size: A Tensor of type float32. A scalar float tensor. Any box that has a smaller size than min_size will be discarded.
  • post_nms_topn: An optional int. Defaults to 300. An integer. Maximum number of rois in the output.
  • name: A name for the operation (optional).

Returns:

A tuple of Tensor objects (rois, roi_probabilities).

  • rois: A Tensor of type float32.
  • roi_probabilities: A Tensor of type float32.