def _get_all_ious(bbox_coords, conv_rows, conv_cols, anchor_dims, img_width,
img_height, stride):
# not used anymore, might be useful to keep around as a reference
num_boxes = conv_rows * conv_cols * len(anchor_dims)
num_gt_boxes = len(bbox_coords)
result = np.zeros((num_boxes, num_gt_boxes))
out_of_bounds_idxs = []
num_boxes = conv_rows * conv_cols * len(anchor_dims)
for i in range(num_boxes):
y, x, anchor_idx = _idx_to_conv(i, conv_cols, len(anchor_dims))
x_center, y_center = _get_conv_center(x, y, stride)
anchor_height, anchor_width = anchor_dims[anchor_idx]
anchor_box = Box.from_center_dims_int(x_center, y_center, anchor_width,
anchor_height)
if _out_of_bounds(anchor_box, img_width, img_height):
out_of_bounds_idxs.append(i)
continue
for bbox_idx in range(num_gt_boxes):
iou = calc_iou(bbox_coords[bbox_idx], anchor_box.corners)
result[i, bbox_idx] = iou
return result, out_of_bounds_idxs
def _process(self, image):
# internal method, performs the expensive calculations needed to produce training inputs.
conv_rows, conv_cols = self.calc_conv_dims(image.height, image.width)
num_anchors = conv_rows * conv_cols * len(self.anchor_dims)
bbreg_targets = np.zeros((num_anchors, 4), dtype=np.float32)
can_use = np.zeros(num_anchors, dtype=np.bool)
is_pos = np.zeros(num_anchors, dtype=np.bool)
gt_box_coords = get_bbox_coords(image.gt_boxes)
anchor_coords = _get_all_anchor_coords(conv_rows, conv_cols,
self.anchor_dims, self.stride)
out_of_bounds_idxs = _get_out_of_bounds_idxs(anchor_coords,
image.width, image.height)
all_ious = cross_ious(anchor_coords, gt_box_coords)
# all_ious, out_of_bounds_idxs = get_all_ious_faster(gt_box_coords, conv_rows, conv_cols, ANCHORS_PER_LOC, image.width, image.height, self.stride)
max_iou_by_anchor = np.amax(all_ious, axis=1)
max_idx_by_anchor = np.argmax(all_ious, axis=1)
max_iou_by_gt_box = np.amax(all_ious, axis=0)
max_idx_by_gt_box = np.argmax(all_ious, axis=0)
# anchors with more than 0.7 IOU with a gt box are positives
pos_box_idxs = np.where(max_iou_by_anchor > POS_OVERLAP)[0]
# for each gt box, the highest non-zero IOU anchor is a positive
eligible_idxs = np.where(max_iou_by_gt_box > 0.0)
more_pos_box_idxs = max_idx_by_gt_box[eligible_idxs]
total_pos_idxs = np.unique(
np.concatenate((pos_box_idxs, more_pos_box_idxs)))
can_use[total_pos_idxs] = 1
is_pos[total_pos_idxs] = 1
# don't bother optimizing, profiling showed this loop's runtime is negligible
for box_idx in total_pos_idxs:
y, x, anchor_idx = _idx_to_conv(box_idx, conv_cols,
len(self.anchor_dims))
x_center, y_center = _get_conv_center(x, y, self.stride)
anchor_height, anchor_width = self.anchor_dims[anchor_idx]
anchor_box = Box.from_center_dims_int(x_center, y_center,
anchor_width, anchor_height)
gt_box_idx = max_idx_by_anchor[box_idx]
reg_params = get_reg_params(anchor_box.corners,
gt_box_coords[gt_box_idx])
bbreg_targets[box_idx, :] = BBREG_MULTIPLIERS * reg_params
neg_box_idxs = np.where(
np.logical_and(is_pos == 0, max_iou_by_anchor < NEG_OVERLAP))[0]
can_use[neg_box_idxs] = 1
can_use[out_of_bounds_idxs] = 0
self._cache[image.cache_key] = {
'can_use': can_use,
'is_pos': is_pos,
'bbreg_targets': bbreg_targets
}
