def _cal_anchor_target(self, label, valid_anchor, gt_bbox, anchor_label):
num_anchor = valid_anchor.shape[0]
reg_target = np.zeros(shape=(num_anchor, 4), dtype=np.float32)
reg_weight = np.zeros(shape=(num_anchor, 4), dtype=np.float32)
fg_index = np.where(label == 1)[0]
if len(fg_index) > 0:
reg_target[fg_index] = bbox_transform(valid_anchor[fg_index], gt_bbox[anchor_label[fg_index], :4])
reg_weight[fg_index, :] = 1.0
return reg_target, reg_weight
def _assign_label_to_anchor(self, valid_anchor, gt_bbox, neg_thr, pos_thr,
min_pos_thr):
num_anchor = valid_anchor.shape[0]
cls_label = np.full(shape=(num_anchor, ),
fill_value=-1,
dtype=np.float32)
reg_target = np.zeros(shape=(num_anchor, 4), dtype=np.float32)
reg_weight = np.zeros(shape=(num_anchor, 4), dtype=np.float32)
if len(gt_bbox) > 0:
# num_anchor x num_gt
overlaps = bbox_overlaps_cython(
valid_anchor.astype(np.float32, copy=False),
gt_bbox.astype(np.float32, copy=False))
max_overlaps = overlaps.max(axis=1)
argmax_overlaps = overlaps.argmax(axis=1)
gt_max_overlaps = overlaps.max(axis=0)
# TODO: speed up this
# TODO: fix potentially assigning wrong anchors as positive
# A correct implementation is given as
# gt_argmax_overlaps = np.where((overlaps.transpose() == gt_max_overlaps[:, None]) &
# (overlaps.transpose() >= min_pos_thr))[1]
gt_argmax_overlaps = np.where((overlaps == gt_max_overlaps)
& (overlaps >= min_pos_thr))
# anchor class
cls_label[max_overlaps < neg_thr] = 0
# fg label: for each gt, anchor with highest overlap
cls_label[gt_argmax_overlaps[0]] = gt_bbox[gt_argmax_overlaps[1],
4]
# fg label: above threshold IoU
cls_label[max_overlaps >= pos_thr] = gt_bbox[argmax_overlaps[
max_overlaps >= pos_thr], 4]
# anchor regression
reg_target[:] = bbox_transform(valid_anchor,
gt_bbox[argmax_overlaps, :4])
reg_weight[cls_label >= 1, :] = 1.0
else:
cls_label[:] = 0
return cls_label, reg_target, reg_weight