def get_anchor_labels(self, anchors, gt_boxes, labels):
# This function will modify labels and return the filtered inds
NA, NB = len(anchors), len(gt_boxes)
assert NB > 0 # empty images should have been filtered already
# ##########
anchor_state = np.zeros((NA, ), dtype='int32') - 1
anchor_labels = np.zeros((NA, ), dtype='int32')
anchor_boxes = np.zeros((NA, 4), dtype='float32')
box_ious = np_iou(anchors, gt_boxes) # NA x NB
# for each anchor box choose the groundtruth box with largest iou, set iou<0.4 as backgroud, ignore 0.4-0.5
max_iou = box_ious.max(axis=1) # NA
positive_anchor_indices = np.where(
max_iou > cfg.ANCHOR.POSITIVE_ANCHOR_THRESH)[0]
negative_anchor_indices = np.where(
max_iou < cfg.ANCHOR.NEGATIVE_ANCHOR_THRESH)[0]
positive_iou = box_ious[positive_anchor_indices]
matched_gt_box_indices = positive_iou.argmax(axis=1)
anchor_labels[positive_anchor_indices] = labels[matched_gt_box_indices]
anchor_state[positive_anchor_indices] = 1
anchor_boxes[positive_anchor_indices] = gt_boxes[
matched_gt_box_indices]
anchor_state[negative_anchor_indices] = 0
fg_boxes = anchor_boxes[anchor_state == 1]
matched_anchors = anchors[anchor_state == 1]
##select and normlised the box coordinate
fg_boxes[:, 0::2] = fg_boxes[:, 0::2] / self.max_size[1]
fg_boxes[:, 1::2] = fg_boxes[:, 1::2] / self.max_size[0]
matched_anchors[:, 0::2] = matched_anchors[:, 0::2] / self.max_size[1]
matched_anchors[:, 1::2] = matched_anchors[:, 1::2] / self.max_size[0]
encode_fg_boxes = encode(fg_boxes, matched_anchors)
anchor_boxes[anchor_state == 1] = encode_fg_boxes
anchor_labels = np.stack([anchor_labels, anchor_state])
#
return anchor_labels, anchor_boxes
def get_anchor_labels(self, anchors, gt_boxes, labels):
# This function will modify labels and return the filtered inds
NA, NB = len(anchors), len(gt_boxes)
assert NB > 0 # empty images should have been filtered already
# ##########
anchor_matched_already = np.zeros((NA, ), dtype='int32')
gt_boxes_mathed_already = np.zeros((NB, ), dtype='int32')
anchor_labels = np.zeros((NA, ), dtype='int32')
anchor_boxes = np.zeros((NA, 4), dtype='float32')
box_ious = np_iou(anchors, gt_boxes) # NA x NB
# for each anchor box choose the groundtruth box with largest iou
max_iou = box_ious.max(axis=1) # NA
positive_anchor_indices = np.where(
max_iou > cfg.ANCHOR.POSITIVE_ANCHOR_THRESH)[0]
# negative_anchor_indices = np.where(max_iou < cfg.ANCHOR.NEGATIVE_ANCHOR_THRESH)[0]
positive_iou = box_ious[positive_anchor_indices]
matched_gt_box_indices = positive_iou.argmax(axis=1)
anchor_labels[positive_anchor_indices] = labels[matched_gt_box_indices]
anchor_boxes[positive_anchor_indices] = gt_boxes[
matched_gt_box_indices]
anchor_matched_already[
positive_anchor_indices] = 1 #### marked as matched
gt_boxes_mathed_already[
matched_gt_box_indices] = 1 #### marked as matched
if np.sum(anchor_matched_already) > 0:
n = np.sum(anchor_matched_already) / np.sum(
gt_boxes_mathed_already)
else:
n = cfg.ANCHOR.AVG_MATCHES
n = n if n > cfg.ANCHOR.AVG_MATCHES else cfg.ANCHOR.AVG_MATCHES
if not cfg.ANCHOR.super_match:
n = cfg.ANCHOR.AVG_MATCHES
# some gt_boxes may not matched, find them and match them with n anchors for each gt box
box_ious[box_ious < cfg.ANCHOR.NEGATIVE_ANCHOR_THRESH] = 0
sorted_ious = np.argsort(-box_ious, axis=0)
sorted_ious = sorted_ious[np.logical_not(anchor_matched_already)]
for i in range(0, len(gt_boxes_mathed_already)):
matched_count = np.sum(
matched_gt_box_indices == gt_boxes_mathed_already[i])
if matched_count >= n:
continue
else:
for j in range(0, int(n - matched_count)):
if box_ious[sorted_ious[j]
[i]][i] > cfg.ANCHOR.NEGATIVE_ANCHOR_THRESH:
anchor_labels[sorted_ious[j][i]] = labels[i]
anchor_boxes[sorted_ious[j][i]] = gt_boxes[i]
anchor_matched_already[sorted_ious[j][i]] = 1
gt_boxes_mathed_already[i] = 1
fg_boxes = anchor_boxes[anchor_matched_already.astype(np.bool)]
matched_anchors = anchors[anchor_matched_already.astype(np.bool)]
##select and normlised the box coordinate
fg_boxes[:, 0::2] = fg_boxes[:, 0::2] / self.max_size[0]
fg_boxes[:, 1::2] = fg_boxes[:, 1::2] / self.max_size[1]
matched_anchors[:, 0::2] = matched_anchors[:, 0::2] / self.max_size[0]
matched_anchors[:, 1::2] = matched_anchors[:, 1::2] / self.max_size[1]
encode_fg_boxes = encode(fg_boxes, matched_anchors)
anchor_boxes[anchor_matched_already.astype(np.bool)] = encode_fg_boxes
# assert len(fg_inds) + np.sum(anchor_labels == 0) == cfg.ANCHOR.BATCH_PER_IM
return anchor_labels, anchor_boxes