def _sample_output(self, all_rois, gt_boxes, im_scale, gt_masks, mask_info,
init_state0):
overlaps = bbox_overlaps(
np.ascontiguousarray(all_rois[:, 1:5], dtype=np.float),
np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
gt_assignment = overlaps.argmax(axis=1)
max_overlaps = overlaps.max(axis=1)
labels = gt_boxes[gt_assignment, 4]
# Sample foreground indexes
fg_inds = np.where(max_overlaps >= cfg.TRAIN.BBOX_THRESH)[0]
bg_inds = np.where(max_overlaps < cfg.TRAIN.BBOX_THRESH)[0]
keep_inds = np.append(fg_inds, bg_inds).astype(int)
# Select sampled values from various arrays:
labels = labels[keep_inds]
# Clamp labels for the background RoIs to 0
labels[len(fg_inds):] = 0
rois = all_rois[keep_inds]
#init_states = init_state0[keep_inds]
init_states = init_state0
bbox_target_data = bbox_compute_targets(
rois[:, 1:5],
gt_boxes[gt_assignment[keep_inds], :4],
normalize=True)
bbox_target_data = np.hstack((labels[:, np.newaxis], bbox_target_data))\
.astype(np.float32, copy=False)
bbox_targets, bbox_inside_weights = get_bbox_regression_label(
bbox_target_data, 21)
scaled_rois = rois[:, 1:5] / float(im_scale)
scaled_gt_boxes = gt_boxes[:, :4] / float(im_scale)
pos_masks = np.zeros((len(keep_inds), 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
top_mask_info = np.zeros((len(keep_inds), 12))
top_mask_info[len(fg_inds):, :] = -1
for i, val in enumerate(fg_inds):
gt_box = scaled_gt_boxes[gt_assignment[val]]
gt_box = np.around(gt_box).astype(int)
ex_box = np.around(scaled_rois[i]).astype(int)
gt_mask = gt_masks[gt_assignment[val]]
gt_mask_info = mask_info[gt_assignment[val]]
gt_mask = gt_mask[0:gt_mask_info[0], 0:gt_mask_info[1]]
# regression targets is the intersection of bounding box and gt mask
ex_mask = intersect_mask(ex_box, gt_box, gt_mask)
pos_masks[i, ...] = ex_mask
top_mask_info[i, 0] = gt_assignment[val]
top_mask_info[i, 1] = gt_mask_info[0]
top_mask_info[i, 2] = gt_mask_info[1]
top_mask_info[i, 3] = labels[i]
top_mask_info[i, 4:8] = ex_box
top_mask_info[i, 8:12] = gt_box
return labels, rois, fg_inds, keep_inds, pos_masks, top_mask_info, bbox_targets, bbox_inside_weights, init_states
def _sample_output(self, all_rois, gt_boxes, im_scale, gt_masks, mask_info):
overlaps = bbox_overlaps(
np.ascontiguousarray(all_rois[:, 1:5], dtype=np.float),
np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
gt_assignment = overlaps.argmax(axis=1)
max_overlaps = overlaps.max(axis=1)
labels = gt_boxes[gt_assignment, 4]
# Sample foreground indexes
fg_inds = np.where(max_overlaps >= cfg.TRAIN.BBOX_THRESH)[0]
bg_inds = np.where(max_overlaps < cfg.TRAIN.BBOX_THRESH)[0]
keep_inds = np.append(fg_inds, bg_inds).astype(int)
# Select sampled values from various arrays:
labels = labels[keep_inds]
# Clamp labels for the background RoIs to 0
labels[len(fg_inds):] = 0
rois = all_rois[keep_inds]
bbox_target_data = bbox_compute_targets(
rois[:, 1:5], gt_boxes[gt_assignment[keep_inds], :4], normalize=True)
bbox_target_data = np.hstack((labels[:, np.newaxis], bbox_target_data))\
.astype(np.float32, copy=False)
bbox_targets, bbox_inside_weights = get_bbox_regression_label(
bbox_target_data, 21)
scaled_rois = rois[:, 1:5] / float(im_scale)
scaled_gt_boxes = gt_boxes[:, :4] / float(im_scale)
pos_masks = np.zeros((len(keep_inds), 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
top_mask_info = np.zeros((len(keep_inds), 12))
top_mask_info[len(fg_inds):, :] = -1
for i, val in enumerate(fg_inds):
gt_box = scaled_gt_boxes[gt_assignment[val]]
gt_box = np.around(gt_box).astype(int)
ex_box = np.around(scaled_rois[i]).astype(int)
gt_mask = gt_masks[gt_assignment[val]]
gt_mask_info = mask_info[gt_assignment[val]]
gt_mask = gt_mask[0:gt_mask_info[0], 0:gt_mask_info[1]]
# regression targets is the intersection of bounding box and gt mask
ex_mask = intersect_mask(ex_box, gt_box, gt_mask)
pos_masks[i, ...] = ex_mask
top_mask_info[i, 0] = gt_assignment[val]
top_mask_info[i, 1] = gt_mask_info[0]
top_mask_info[i, 2] = gt_mask_info[1]
top_mask_info[i, 3] = labels[i]
top_mask_info[i, 4:8] = ex_box
top_mask_info[i, 8:12] = gt_box
return labels, rois, fg_inds, keep_inds, pos_masks, top_mask_info, bbox_targets, bbox_inside_weights
def _sample_rois(all_rois, gt_boxes, rois_per_image, num_classes, gt_masks, im_scale, mask_info):
"""
Generate a random sample of RoIs comprising
foreground and background examples.
"""
# overlaps: (rois x gt_boxes)
overlaps = bbox_overlaps(
np.ascontiguousarray(all_rois[:, 1:5], dtype=np.float),
np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
gt_assignment = overlaps.argmax(axis=1)
max_overlaps = overlaps.max(axis=1)
labels = gt_boxes[gt_assignment, 4]
# Sample foreground indexes
fg_inds = []
for i in xrange(len(cfg.TRAIN.FG_FRACTION)):
cur_inds = np.where((max_overlaps >= cfg.TRAIN.FG_THRESH_LO[i]) &
(max_overlaps <= cfg.TRAIN.FG_THRESH_HI[i]))[0]
cur_rois_this_image = min(cur_inds.size, np.round(rois_per_image *
cfg.TRAIN.FG_FRACTION[i]))
if cur_inds.size > 0:
cur_inds = npr.choice(cur_inds, size=cur_rois_this_image, replace=False)
fg_inds = np.hstack((fg_inds, cur_inds))
fg_inds = np.unique(fg_inds)
fg_rois_per_image = fg_inds.size
# Sample background indexes according to number of foreground
bg_rois_per_this_image = rois_per_image - fg_rois_per_image
bg_inds = []
for i in xrange(len(cfg.TRAIN.BG_FRACTION)):
cur_inds = np.where((max_overlaps >= cfg.TRAIN.BG_THRESH_LO[i]) &
(max_overlaps <= cfg.TRAIN.BG_THRESH_HI[i]))[0]
cur_rois_this_image = min(cur_inds.size, np.round(bg_rois_per_this_image *
cfg.TRAIN.BG_FRACTION[i]))
if cur_inds.size > 0:
cur_inds = npr.choice(cur_inds, size=cur_rois_this_image, replace=False)
bg_inds = np.hstack((bg_inds, cur_inds))
bg_inds = np.unique(bg_inds)
# The indices that we're selecting (both fg and bg)
keep_inds = np.append(fg_inds, bg_inds).astype(int)
# Select sampled values from various arrays:
labels = labels[keep_inds]
# Clamp labels for the background RoIs to 0
labels[fg_rois_per_image:] = 0
rois = all_rois[keep_inds]
bbox_target_data = bbox_compute_targets(
rois[:, 1:5], gt_boxes[gt_assignment[keep_inds], :4], normalize=True)
bbox_target_data = np.hstack((labels[:, np.newaxis], bbox_target_data))\
.astype(np.float32, copy=False)
bbox_targets, bbox_inside_weights = get_bbox_regression_label(
bbox_target_data, num_classes)
bbox_outside_weights = np.array(bbox_inside_weights > 0).astype(np.float32)
blobs = {
'rois': rois,
'labels': labels,
'bbox_targets': bbox_targets,
'bbox_inside_weights': bbox_inside_weights,
'bbox_outside_weights': bbox_outside_weights
}
if cfg.MNC_MODE:
scaled_rois = rois[:, 1:5] / float(im_scale)
# map to original image space
scaled_gt_boxes = gt_boxes[:, :4] / float(im_scale)
pos_masks = np.zeros((len(keep_inds), 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
top_mask_info = np.zeros((len(keep_inds), 12))
top_mask_info[len(fg_inds):, :] = -1
for i, val in enumerate(fg_inds):
gt_box = scaled_gt_boxes[gt_assignment[val]]
gt_box = np.around(gt_box).astype(int)
ex_box = np.around(scaled_rois[i]).astype(int)
gt_mask = gt_masks[gt_assignment[val]]
gt_mask_info = mask_info[gt_assignment[val]]
gt_mask = gt_mask[0:gt_mask_info[0], 0:gt_mask_info[1]]
# calculate mask regression targets
# (intersection of bounding box and gt mask)
ex_mask = intersect_mask(ex_box, gt_box, gt_mask)
pos_masks[i, ...] = ex_mask
top_mask_info[i, 0] = gt_assignment[val]
top_mask_info[i, 1] = gt_mask_info[0]
top_mask_info[i, 2] = gt_mask_info[1]
top_mask_info[i, 3] = labels[i]
top_mask_info[i, 4:8] = ex_box
top_mask_info[i, 8:12] = gt_box
mask_weight = np.zeros((rois.shape[0], 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
# only assign box-level foreground as positive mask regression
mask_weight[0:len(fg_inds), :, :, :] = 1
blobs['mask_targets'] = pos_masks
blobs['mask_weight'] = mask_weight
blobs['gt_masks_info'] = top_mask_info
return blobs, fg_inds, bg_inds, keep_inds
def process_roidb(file_start, file_end, db):
for cnt in xrange(file_start, file_end):
f = file_list[cnt]
full_file = os.path.join(input_dir, f)
output_cache = os.path.join(output_dir, f.split('.')[0] + '.mat')
timer_tic = time.time()
if os.path.exists(output_cache):
continue
mcg_mat = sio.loadmat(full_file)
mcg_mask_label = mcg_mat['labels']
mcg_superpixels = mcg_mat['superpixels']
num_proposal = len(mcg_mask_label)
mcg_boxes = np.zeros((num_proposal, 4))
mcg_masks = np.zeros((num_proposal, mask_size, mask_size), dtype=np.bool)
for ind_proposal in xrange(num_proposal):
label = mcg_mask_label[ind_proposal][0][0]
proposal = np.in1d(mcg_superpixels, label).reshape(mcg_superpixels.shape)
[r, c] = np.where(proposal == 1)
y1 = np.min(r)
x1 = np.min(c)
y2 = np.max(r)
x2 = np.max(c)
box = np.array([x1, y1, x2, y2])
proposal = proposal[y1:y2+1, x1:x2+1]
proposal = cv2.resize(proposal.astype(np.float), (mask_size, mask_size), interpolation=cv2.INTER_NEAREST)
mcg_masks[ind_proposal, :, :] = proposal
mcg_boxes[ind_proposal, :] = box
if top_k != -1:
mcg_boxes = mcg_boxes[:top_k, :]
mcg_masks = mcg_masks[:top_k, :]
if db == 'val':
# if we prepare validation data, we only need its masks and boxes
roidb = {
'masks': (mcg_masks >= cfg.BINARIZE_THRESH).astype(bool),
'boxes': mcg_boxes
}
sio.savemat(output_cache, roidb)
use_time = time.time() - timer_tic
print '%d/%d use time %f' % (cnt, len(file_list), use_time)
else:
# Otherwise we need to prepare other information like overlaps
num_mcg = mcg_boxes.shape[0]
gt_roidb = gt_roidbs[cnt]
gt_maskdb = gt_maskdbs[cnt]
gt_boxes = gt_roidb['boxes']
gt_masks = gt_maskdb['gt_masks']
gt_classes = gt_roidb['gt_classes']
num_gt = gt_boxes.shape[0]
num_all = num_gt + num_mcg
# define output structure
det_overlaps = np.zeros((num_all, 1))
seg_overlaps = np.zeros((num_all, 1))
seg_assignment = np.zeros((num_all, 1))
mask_targets = np.zeros((num_all, mask_size, mask_size))
# ------------------------------------------------------
all_boxes = np.vstack((gt_boxes[:, :4], mcg_boxes)).astype(int)
all_masks = np.zeros((num_all, mask_size, mask_size))
for i in xrange(num_gt):
all_masks[i, :, :] = (cv2.resize(gt_masks[i].astype(np.float),
(mask_size, mask_size)))
assert all_masks[num_gt:, :, :].shape == mcg_masks.shape
all_masks[num_gt:, :, :] = mcg_masks
# record bounding box overlaps
cur_overlap = bbox_overlaps(all_boxes.astype(np.float), gt_boxes.astype(np.float))
seg_assignment = cur_overlap.argmax(axis=1)
det_overlaps = cur_overlap.max(axis=1)
seg_assignment[det_overlaps == 0] = -1
# record mask region overlaps
seg_overlaps[:num_gt] = 1.0
for i in xrange(num_gt, num_all):
cur_mask = cv2.resize(all_masks[i, :, :].astype(np.float),
(all_boxes[i, 2] - all_boxes[i, 0] + 1,
all_boxes[i, 3] - all_boxes[i, 1] + 1)) >= cfg.BINARIZE_THRESH
for mask_ind in xrange(len(gt_masks)):
gt_mask = gt_masks[mask_ind]
gt_roi = gt_roidb['boxes'][mask_ind]
cur_ov = mask_overlap(all_boxes[i, :], gt_roi, cur_mask, gt_mask)
seg_overlaps[i] = max(seg_overlaps[i], cur_ov)
output_label = np.zeros((num_all, 1))
for i in xrange(num_all):
if seg_assignment[i] == -1:
continue
cur_ind = seg_assignment[i]
output_label[i] = gt_classes[seg_assignment[i]]
mask_targets[i, :, :] = intersect_mask(all_boxes[i, :], gt_roidb['boxes'][cur_ind], gt_masks[cur_ind])
# Some of the array need to insert a new axis to be consistent of savemat method
roidb = {
'masks': (all_masks >= cfg.BINARIZE_THRESH).astype(bool),
'boxes': all_boxes,
'det_overlap': det_overlaps[:, np.newaxis],
'seg_overlap': seg_overlaps,
'mask_targets': (mask_targets >= cfg.BINARIZE_THRESH).astype(bool),
'gt_classes': gt_classes[:, np.newaxis],
'output_label': output_label,
'gt_assignment': seg_assignment[:, np.newaxis],
'Flip': False
}
sio.savemat(output_cache, roidb)
use_time = time.time() - timer_tic
print '%d/%d use time %f' % (cnt, len(file_list), use_time)
def process_roidb(file_start, file_end, db):
for cnt in xrange(file_start, file_end):
f = file_list[cnt]
full_file = os.path.join(input_dir, f)
output_cache = os.path.join(output_dir, f.split('.')[0] + '.mat')
timer_tic = time.time()
if os.path.exists(output_cache):
continue
mcg_mat = sio.loadmat(full_file)
mcg_mask_label = mcg_mat['labels']
mcg_superpixels = mcg_mat['superpixels']
num_proposal = len(mcg_mask_label)
mcg_boxes = np.zeros((num_proposal, 4))
mcg_masks = np.zeros((num_proposal, mask_size, mask_size),
dtype=np.bool)
for ind_proposal in xrange(num_proposal):
label = mcg_mask_label[ind_proposal][0][0]
proposal = np.in1d(mcg_superpixels,
label).reshape(mcg_superpixels.shape)
[r, c] = np.where(proposal == 1)
y1 = np.min(r)
x1 = np.min(c)
y2 = np.max(r)
x2 = np.max(c)
box = np.array([x1, y1, x2, y2])
proposal = proposal[y1:y2 + 1, x1:x2 + 1]
proposal = cv2.resize(proposal.astype(np.float),
(mask_size, mask_size),
interpolation=cv2.INTER_NEAREST)
mcg_masks[ind_proposal, :, :] = proposal
mcg_boxes[ind_proposal, :] = box
if top_k != -1:
mcg_boxes = mcg_boxes[:top_k, :]
mcg_masks = mcg_masks[:top_k, :]
if db == 'val':
# if we prepare validation data, we only need its masks and boxes
roidb = {
'masks': (mcg_masks >= cfg.BINARIZE_THRESH).astype(bool),
'boxes': mcg_boxes
}
sio.savemat(output_cache, roidb)
use_time = time.time() - timer_tic
print '%d/%d use time %f' % (cnt, len(file_list), use_time)
else:
# Otherwise we need to prepare other information like overlaps
num_mcg = mcg_boxes.shape[0]
gt_roidb = gt_roidbs[cnt]
gt_maskdb = gt_maskdbs[cnt]
gt_boxes = gt_roidb['boxes']
gt_masks = gt_maskdb['gt_masks']
gt_classes = gt_roidb['gt_classes']
num_gt = gt_boxes.shape[0]
num_all = num_gt + num_mcg
# define output structure
det_overlaps = np.zeros((num_all, 1))
seg_overlaps = np.zeros((num_all, 1))
seg_assignment = np.zeros((num_all, 1))
mask_targets = np.zeros((num_all, mask_size, mask_size))
# ------------------------------------------------------
all_boxes = np.vstack((gt_boxes[:, :4], mcg_boxes)).astype(int)
all_masks = np.zeros((num_all, mask_size, mask_size))
for i in xrange(num_gt):
all_masks[i, :, :] = (cv2.resize(gt_masks[i].astype(np.float),
(mask_size, mask_size)))
assert all_masks[num_gt:, :, :].shape == mcg_masks.shape
all_masks[num_gt:, :, :] = mcg_masks
# record bounding box overlaps
cur_overlap = bbox_overlaps(all_boxes.astype(np.float),
gt_boxes.astype(np.float))
seg_assignment = cur_overlap.argmax(axis=1)
det_overlaps = cur_overlap.max(axis=1)
seg_assignment[det_overlaps == 0] = -1
# record mask region overlaps
seg_overlaps[:num_gt] = 1.0
for i in xrange(num_gt, num_all):
cur_mask = cv2.resize(
all_masks[i, :, :].astype(np.float),
(all_boxes[i, 2] - all_boxes[i, 0] + 1, all_boxes[i, 3] -
all_boxes[i, 1] + 1)) >= cfg.BINARIZE_THRESH
for mask_ind in xrange(len(gt_masks)):
gt_mask = gt_masks[mask_ind]
gt_roi = gt_roidb['boxes'][mask_ind]
cur_ov = mask_overlap(all_boxes[i, :], gt_roi, cur_mask,
gt_mask)
seg_overlaps[i] = max(seg_overlaps[i], cur_ov)
output_label = np.zeros((num_all, 1))
for i in xrange(num_all):
if seg_assignment[i] == -1:
continue
cur_ind = seg_assignment[i]
output_label[i] = gt_classes[seg_assignment[i]]
mask_targets[i, :, :] = intersect_mask(
all_boxes[i, :], gt_roidb['boxes'][cur_ind],
gt_masks[cur_ind])
# Some of the array need to insert a new axis to be consistent of savemat method
roidb = {
'masks': (all_masks >= cfg.BINARIZE_THRESH).astype(bool),
'boxes': all_boxes,
'det_overlap': det_overlaps[:, np.newaxis],
'seg_overlap': seg_overlaps,
'mask_targets':
(mask_targets >= cfg.BINARIZE_THRESH).astype(bool),
'gt_classes': gt_classes[:, np.newaxis],
'output_label': output_label,
'gt_assignment': seg_assignment[:, np.newaxis],
'Flip': False
}
sio.savemat(output_cache, roidb)
use_time = time.time() - timer_tic
print '%d/%d use time %f' % (cnt, len(file_list), use_time)
def _sample_rois(all_rois, gt_boxes, rois_per_image, num_classes, gt_masks,
im_scale, mask_info):
"""
Generate a random sample of RoIs comprising
foreground and background examples.
"""
# overlaps: (rois x gt_boxes)
overlaps = bbox_overlaps(
np.ascontiguousarray(all_rois[:, 1:5], dtype=np.float),
np.ascontiguousarray(gt_boxes[:, :4], dtype=np.float))
gt_assignment = overlaps.argmax(axis=1)
max_overlaps = overlaps.max(axis=1)
labels = gt_boxes[gt_assignment, 4]
# Sample foreground indexes
fg_inds = []
for i in xrange(len(cfg.TRAIN.FG_FRACTION)):
cur_inds = np.where((max_overlaps >= cfg.TRAIN.FG_THRESH_LO[i])
& (max_overlaps <= cfg.TRAIN.FG_THRESH_HI[i]))[0]
cur_rois_this_image = min(
cur_inds.size, np.round(rois_per_image * cfg.TRAIN.FG_FRACTION[i]))
if cur_inds.size > 0:
cur_inds = npr.choice(cur_inds,
size=cur_rois_this_image,
replace=False)
fg_inds = np.hstack((fg_inds, cur_inds))
fg_inds = np.unique(fg_inds)
fg_rois_per_image = fg_inds.size
# Sample background indexes according to number of foreground
bg_rois_per_this_image = rois_per_image - fg_rois_per_image
bg_inds = []
for i in xrange(len(cfg.TRAIN.BG_FRACTION)):
cur_inds = np.where((max_overlaps >= cfg.TRAIN.BG_THRESH_LO[i])
& (max_overlaps <= cfg.TRAIN.BG_THRESH_HI[i]))[0]
cur_rois_this_image = min(
cur_inds.size,
np.round(bg_rois_per_this_image * cfg.TRAIN.BG_FRACTION[i]))
if cur_inds.size > 0:
cur_inds = npr.choice(cur_inds,
size=cur_rois_this_image,
replace=False)
bg_inds = np.hstack((bg_inds, cur_inds))
bg_inds = np.unique(bg_inds)
# The indices that we're selecting (both fg and bg)
keep_inds = np.append(fg_inds, bg_inds).astype(int)
# Select sampled values from various arrays:
labels = labels[keep_inds]
# Clamp labels for the background RoIs to 0
labels[fg_rois_per_image:] = 0
rois = all_rois[keep_inds]
bbox_target_data = bbox_compute_targets(
rois[:, 1:5], gt_boxes[gt_assignment[keep_inds], :4], normalize=True)
bbox_target_data = np.hstack((labels[:, np.newaxis], bbox_target_data))\
.astype(np.float32, copy=False)
bbox_targets, bbox_inside_weights = get_bbox_regression_label(
bbox_target_data, num_classes)
bbox_outside_weights = np.array(bbox_inside_weights > 0).astype(np.float32)
blobs = {
'rois': rois,
'labels': labels,
'bbox_targets': bbox_targets,
'bbox_inside_weights': bbox_inside_weights,
'bbox_outside_weights': bbox_outside_weights
}
if cfg.MNC_MODE:
scaled_rois = rois[:, 1:5] / float(im_scale)
# map to original image space
scaled_gt_boxes = gt_boxes[:, :4] / float(im_scale)
pos_masks = np.zeros((len(keep_inds), 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
top_mask_info = np.zeros((len(keep_inds), 12))
top_mask_info[len(fg_inds):, :] = -1
for i, val in enumerate(fg_inds):
gt_box = scaled_gt_boxes[gt_assignment[val]]
gt_box = np.around(gt_box).astype(int)
ex_box = np.around(scaled_rois[i]).astype(int)
gt_mask = gt_masks[gt_assignment[val]]
gt_mask_info = mask_info[gt_assignment[val]]
gt_mask = gt_mask[0:gt_mask_info[0], 0:gt_mask_info[1]]
# calculate mask regression targets
# (intersection of bounding box and gt mask)
ex_mask = intersect_mask(ex_box, gt_box, gt_mask)
pos_masks[i, ...] = ex_mask
top_mask_info[i, 0] = gt_assignment[val]
top_mask_info[i, 1] = gt_mask_info[0]
top_mask_info[i, 2] = gt_mask_info[1]
top_mask_info[i, 3] = labels[i]
top_mask_info[i, 4:8] = ex_box
top_mask_info[i, 8:12] = gt_box
mask_weight = np.zeros(
(rois.shape[0], 1, cfg.MASK_SIZE, cfg.MASK_SIZE))
# only assign box-level foreground as positive mask regression
mask_weight[0:len(fg_inds), :, :, :] = 1
blobs['mask_targets'] = pos_masks
blobs['mask_weight'] = mask_weight
blobs['gt_masks_info'] = top_mask_info
return blobs, fg_inds, bg_inds, keep_inds