def build_rpn_targets(image_shape, anchors, gt_class_ids, gt_boxes, config):
# RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
# RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))]
rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4))
crowd_ix = np.where(gt_class_ids < 0)[0]
if crowd_ix.shape[0] > 0:
# Filter out crowds from ground truth class IDs and boxes
non_crowd_ix = np.where(gt_class_ids > 0)[0]
crowd_boxes = gt_boxes[crowd_ix]
gt_class_ids = gt_class_ids[non_crowd_ix]
gt_boxes = gt_boxes[non_crowd_ix]
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes)
crowd_iou_max = np.amax(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
else:
# All anchors don't intersect a crowd
no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool)
# Compute overlaps [num_anchors, num_gt_boxes]
overlaps = utils.compute_overlaps(anchors, gt_boxes)
anchor_iou_argmax = np.argmax(overlaps, axis=1)
anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
rpn_match[(anchor_iou_max < 0.3) & (no_crowd_bool)] = -1
# 2. Set an anchor for each GT box (regardless of IoU value).
# TODO: If multiple anchors have the same IoU match all of them
gt_iou_argmax = np.argmax(overlaps, axis=0)
rpn_match[gt_iou_argmax] = 1
# 3. Set anchors with high overlap as positive.
rpn_match[anchor_iou_max >= 0.7] = 1
# Subsample to balance positive and negative anchors
# Don't let positives be more than half the anchors
ids = np.where(rpn_match == 1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2)
if extra > 0:
# Reset the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# Same for negative proposals
ids = np.where(rpn_match == -1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE -
np.sum(rpn_match == 1))
if extra > 0:
# Rest the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# For positive anchors, compute shift and scale needed to transform them
# to match the corresponding GT boxes.
ids = np.where(rpn_match == 1)[0]
ix = 0 # index into rpn_bbox
# TODO: use box_refinment() rather than duplicating the code here
for i, a in zip(ids, anchors[ids]):
# Closest gt box (it might have IoU < 0.7)
gt = gt_boxes[anchor_iou_argmax[i]]
# Convert coordinates to center plus width/height.
# GT Box
gt_h = gt[2] - gt[0]
gt_w = gt[3] - gt[1]
gt_center_y = gt[0] + 0.5 * gt_h
gt_center_x = gt[1] + 0.5 * gt_w
# Anchor
a_h = a[2] - a[0]
a_w = a[3] - a[1]
a_center_y = a[0] + 0.5 * a_h
a_center_x = a[1] + 0.5 * a_w
# Compute the bbox refinement that the RPN should predict.
rpn_bbox[ix] = [
(gt_center_y - a_center_y) / a_h,
(gt_center_x - a_center_x) / a_w,
np.log(gt_h / a_h),
np.log(gt_w / a_w),
]
# Normalize
rpn_bbox[ix] /= config.RPN_BBOX_STD_DEV
ix += 1
return rpn_match, rpn_bbox
def build_rpn_targets(image_shape, anchors, gt_class_ids, gt_boxes, config):
"""Given the anchors and GT boxes, compute overlaps and identify positive
anchors and deltas to refine them to match their corresponding GT boxes.
anchors: [num_anchors, (y1, x1, y2, x2)]
gt_class_ids: [num_gt_boxes] Integer class IDs.
gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)]
Returns:
rpn_match: [N] (int32) matches between anchors and GT boxes.
1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas.
"""
# RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
# RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))]
rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4))
# Handle COCO crowds
# A crowd box in COCO is a bounding box around several instances. Exclude
# them from training. A crowd box is given a negative class ID.
crowd_ix = np.where(gt_class_ids < 0)[0]
if crowd_ix.shape[0] > 0:
# Filter out crowds from ground truth class IDs and boxes
non_crowd_ix = np.where(gt_class_ids > 0)[0]
crowd_boxes = gt_boxes[crowd_ix]
gt_class_ids = gt_class_ids[non_crowd_ix]
gt_boxes = gt_boxes[non_crowd_ix]
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes)
crowd_iou_max = np.amax(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
else:
# All anchors don't intersect a crowd
no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool)
# Compute overlaps [num_anchors, num_gt_boxes]
overlaps = utils.compute_overlaps(anchors, gt_boxes)
# Match anchors to GT Boxes
# If an anchor overlaps a GT box with IoU >= 0.7 then it's positive.
# If an anchor overlaps a GT box with IoU < 0.3 then it's negative.
# Neutral anchors are those that don't match the conditions above,
# and they don't influence the loss function.
# However, don't keep any GT box unmatched (rare, but happens). Instead,
# match it to the closest anchor (even if its max IoU is < 0.3).
#
# 1. Set negative anchors first. They get overwritten below if a GT box is
# matched to them. Skip boxes in crowd areas.
anchor_iou_argmax = np.argmax(overlaps, axis=1)
anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
rpn_match[(anchor_iou_max < 0.3) & (no_crowd_bool)] = -1
# 2. Set an anchor for each GT box (regardless of IoU value).
# TODO: If multiple anchors have the same IoU match all of them
gt_iou_argmax = np.argmax(overlaps, axis=0)
rpn_match[gt_iou_argmax] = 1
# 3. Set anchors with high overlap as positive.
rpn_match[anchor_iou_max >= 0.7] = 1
# Subsample to balance positive and negative anchors
# Don't let positives be more than half the anchors
ids = np.where(rpn_match == 1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2)
if extra > 0:
# Reset the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# Same for negative proposals
ids = np.where(rpn_match == -1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE -
np.sum(rpn_match == 1))
if extra > 0:
# Rest the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# For positive anchors, compute shift and scale needed to transform them
# to match the corresponding GT boxes.
ids = np.where(rpn_match == 1)[0]
ix = 0 # index into rpn_bbox
# TODO: use box_refinement() rather than duplicating the code here
for i, a in zip(ids, anchors[ids]):
# Closest gt box (it might have IoU < 0.7)
gt = gt_boxes[anchor_iou_argmax[i]]
# Convert coordinates to center plus width/height.
# GT Box
gt_h = gt[2] - gt[0]
gt_w = gt[3] - gt[1]
gt_center_y = gt[0] + 0.5 * gt_h
gt_center_x = gt[1] + 0.5 * gt_w
# Anchor
a_h = a[2] - a[0]
a_w = a[3] - a[1]
a_center_y = a[0] + 0.5 * a_h
a_center_x = a[1] + 0.5 * a_w
# Compute the bbox refinement that the RPN should predict.
rpn_bbox[ix] = [
(gt_center_y - a_center_y) / a_h,
(gt_center_x - a_center_x) / a_w,
np.log(gt_h / a_h),
np.log(gt_w / a_w),
]
# Normalize
rpn_bbox[ix] /= config.RPN_BBOX_STD_DEV
ix += 1
return rpn_match, rpn_bbox
def build_rpn_targets(image_shape, anchors, gt_class_ids, gt_boxes, config):
"""Given the anchors and GT boxes, compute overlaps and identify positive
anchors and deltas to refine them to match their corresponding GT boxes.
anchors: [num_anchors, (y1, x1, y2, x2)]
gt_class_ids: [num_gt_boxes] Integer class IDs.
gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)]
Returns:
rpn_match: [N] (int32) matches between anchors and GT boxes.
1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas.
"""
# RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
# RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))]
rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4))
# Handle COCO crowds
# A crowd in COCO is a bounding box around several instances. Exclude
# them for training. A crowd box is given a negative class ID.
crowd_ix = np.where(gt_class_ids < 0)[0]
if crowd_ix.shape[0] > 0:
# Filter out crowds from ground truth class IDs and boxes
non_crowd_ix = np.where(gt_class_ids > 0)[0]
crowd_boxes = gt_boxes[crowd_ix]
gt_class_ids = gt_class_ids[non_crowd_ix]
gt_boxes = gt_boxes[non_crowd_ix]
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes)
crowd_iou = np.amax(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
else:
# All anchors don't intersect a crowd
no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool)
# Compute overlaps [num_anchors, num_gt_boxes]
overlaps = utils.compute_overlaps(anchors, gt_boxes)
# Match anchors to GT Boxes
# If an anchor overlaps a GT box with IoU >= 0.7 then it's positive.
# If an anchor overlaps a GT box with IoU < 0.3 then it's negative.
# Neutral anchors are those that don't match the conditions above,
# and they don't influence the loss function.
# However, don't keep any GT box unmatched (rare, but happens). Instead
# match it to the closest anchor ( even if its max IoU is < 0.3).
#
# 1. Set negative anchors first. They get overwritten below if a GT box
# is matched to them. Skip boxes in crowd areas.
anchor_iou_argmax = np.argmax(overlaps, axis=1)
anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
rpn_match[(anchor_ioud_max < 0.3) * (no_crowd_bool)] = -1
# 2. Set an anchor for each GT box (regardless of IoU value).
# If multiple anchors have the same IoU match all of them
gt_ioud_argmax = np.argwhere(overlaps == np.max(overlaps, axis=0))[:, 0]
rpn_match[gt_iou_argmax] = 1
# 3. Set anchors with high overlap as positive.
rpn_match[anchor_iou_max >= 0.7] = 1
# Subsample to balance positive and negative anchors
# Don't let positives be more than hals the anchors
ids = np.where(rpn_match == 1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE -
np.sum(rpn_match == 1))
def test(model):
"""Test the model."""
save_path = "./results/" + args.weights[27:] + "/"
if not os.path.exists(save_path):
os.makedirs(save_path)
per_class_ious = []
box_iou = []
detect_time = 0
for i in range(args.limit, 131):
image = np.load(args.data + "image_np/liver_" + str(i) + ".npy").astype(np.float32) # [H, W, D]
label = np.load(args.data + "label_np/liver_label_" + str(i) + ".npy").astype(np.int32) # [H, W, D, num_class]
gt_bbox = utils.extract_bboxes(label)
gt_bbox = utils.extend_bbox(gt_bbox, label.shape)
nib_label = nib.load("/media/disk1/LiTS/labelTr/segmentation-" + str(i) + ".nii.gz")
affine = nib_label.affine # prepared to save the predicted mask later
ori_shape = nib_label.shape
try:
start_time = time.time()
result = model.detect([image])[0]
detect_time += time.time() - start_time
print("detect_time:", time.time() - start_time)
"""The shape of result: a dict containing
{
"rois": final_rois, [N, (y1, x1, z1, y2, x2, z2)] in real coordinates
"class_ids": final_class_ids, [N]
"scores": final_scores, [N]
"mask": final_mask, [mask_shape[0], mask_shape[1], mask_shape[2]]
}"""
rois = result["rois"]
class_ids = result["class_ids"]
scores = result["scores"]
mask = result["mask"]
if args.stage == 'beginning':
mask = np.zeros(mask.shape).astype(np.int32)
rois = rois.clip(min=0)
rois[:, 3] = rois[:, 3].clip(max=mask.shape[0] - 1)
rois[:, 4] = rois[:, 4].clip(max=mask.shape[1] - 1)
rois[:, 5] = rois[:, 5].clip(max=mask.shape[2] - 1)
rois = rois.astype(np.int32)
# import pdb
# pdb.set_trace()
# compute bbox iou
print("gt_bbox:", gt_bbox, "pred_bbox:", rois)
box_iou.append(utils.compute_overlaps(np.array([gt_bbox]), rois)[0, 0])
if args.stage != 'beginning':
# Prepare the gt-masks and pred-masks to calculate the ious. [H, W, D, num_classes - 1]
gt_masks = np.zeros(label.shape[:3] + (model.config.NUM_CLASSES - 1,))
pred_masks = np.zeros(image.shape + (model.config.NUM_CLASSES - 1,))
for j in range(model.config.NUM_CLASSES - 1):
gt_masks[:, :, :, j][label == j + 1] = 1
pred_masks[:, :, :, j][mask == j + 1] = 1
# calculate different kind of ious
per_class_iou = utils.compute_per_class_mask_iou(gt_masks, pred_masks)
per_class_ious.append(per_class_iou)
# Save the results
if args.save:
# Draw bboxes
if args.bbox:
for j in range(rois.shape[0]):
y1, x1, z1, y2, x2, z2 = rois[j, :]
mask[y1:y2, x1:x2, z1:z2] = 100
mask = resize(mask, ori_shape, order=0, mode='constant', preserve_range=True, anti_aliasing=False)
vol = nib.Nifti1Image(mask.astype(np.uint8), affine)
if args.stage != 'beginning':
nib.save(vol, save_path + str(per_class_iou.mean()) + "_liver_" + str(i) + ".nii.gz")
print("liver_" + str(i) + " detected done. iou = " + str(per_class_iou))
else:
nib.save(vol, save_path + str(box_iou[-1]) + "_liver_" + str(i) + ".nii.gz")
print("liver_" + str(i) + " detected done. box_iou = " + str(box_iou[-1]))
except:
print("detect error!")
pass
print("Test completed.")
# Print the iou results.
box_iou = np.array(box_iou)
print("box iou:", box_iou)
print("mean:", box_iou.mean())
if args.stage != 'beginning':
per_class_ious = np.array(per_class_ious)
print("per class iou mean:", np.mean(per_class_ious, axis=0))
print("std:", np.std(per_class_ious, axis=0))
print("Total ious mean:", per_class_ious.mean())
print("Total detect time:", detect_time)
def build_rpn_targets(anchors, gt_boxes):
"""Given the anchors and GT boxes, compute overlaps and identify positive
anchors and deltas to refine them to match their corresponding GT boxes.
Args:
anchors: [num_anchors, (y1, x1, y2, x2)]
gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)]
Returns:
rpn_match: [N] (int32) matches between anchors and GT boxes.
1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas.
"""
# RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
# RPN bounding boxes: [max_anchors_per_image, (dy, dx, log(dh), log(dw))]
rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4), dtype=np.float32)
# Compute overlaps [num_anchors, num_gt_boxes]
overlaps = utils.compute_overlaps(anchors, gt_boxes)
# Match anchors to GT Boxes
# If an anchor overlaps a GT box with IoU >= 0.7 then it's positive.
# If an anchor overlaps a GT box with IoU < 0.3 then it's negative.
# Neutral anchors are those that don't match the conditions above,
# and they don't influence the loss function.
# However, don't keep any GT box unmatched (rare, but happens). Instead,
# match it to the closest anchor (even if its max IoU is < 0.3).
# 1. Set negative anchors first. They get overwritten below if a GT box is matched to them.
# 每个 anchor 对应有最大 overlap 的 gt_box_id
anchor_iou_argmax = np.argmax(overlaps, axis=1)
# 每个 anchor 和所有 gt_boxes 的最大 overlap
anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
# 如果 anchor 和 所有 gt_boxes 的 overlap 都小于 0.3, 设置该 anchor 为 neg
rpn_match[(anchor_iou_max < 0.3)] = -1
# 2. Set an anchor for each GT box (regardless of IoU value).
# If multiple anchors have the same IoU match all of them
# 竟然会出现一个 gt_box 和多个 anchor 有相同的 iou, [:, 0] 表示只获取这些 anchor 的 id
# overlaps == np.max(overlaps, axis=0) 会进行 broadcast, 得到和 overlaps 相同 shape 的结果
# In [8]: a=np.array([[1, 0, 3, 2],[2, 1, 0, 1]])
# In [9]: a == np.max(a, axis=0)
# Out[9]:
# array([[False, False, True, True],
# [ True, True, False, False]])
# 得到的是和各个 gt_box 有最大 overlap 的 anchor 的 id,
# NOTE: 一个 gt_box 可能和多个 anchors 有最大 overlap 这些 anchor 都认为是 pos
# 一个 anchor 也可能和多个 gt_boxes 有最大 iou
# argwhere 返回值的 shape 为 (n, len(overlaps.shape)), n 表示符合条件的个数, 第二维表示符合条件的值的下标
gt_iou_argmax = np.argwhere(overlaps == np.max(overlaps, axis=0))[:, 0]
rpn_match[gt_iou_argmax] = 1
# 3. Set anchors with high overlap as positive.
# 如果 anchor 和所有 gt_boxes 的最大的 overlap 大于 0.7, 设置该 anchor 为 pos
rpn_match[anchor_iou_max >= 0.7] = 1
# Subsample to balance positive and negative anchors
# Don't let positives be more than half the anchors
ids = np.where(rpn_match == 1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2)
if extra > 0:
# Reset the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# Same for negative proposals
ids = np.where(rpn_match == -1)[0]
# pos 和 neg anchors 的个数加起来等于 config.RPN_TRAIN_ANCHORS_PER_IMAGE
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE - np.sum(rpn_match == 1))
if extra > 0:
# Rest the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# For positive anchors, compute shift and scale needed to transform them to match the corresponding GT boxes.
ids = np.where(rpn_match == 1)[0]
# index into rpn_bbox
ix = 0
# TODO: use box_refinement() rather than duplicating the code here
for i, a in zip(ids, anchors[ids]):
# Note: 这里的 i 表示的是 pos anchor 在所有 anchors 中的 id
# Closest gt box (it might have IoU < 0.7)
# 取和此 anchor 有最大 iou 的 gt_box
gt = gt_boxes[anchor_iou_argmax[i]]
# Convert coordinates to center plus width/height.
# GT Box
gt_h = gt[2] - gt[0]
gt_w = gt[3] - gt[1]
gt_center_y = gt[0] + 0.5 * gt_h
gt_center_x = gt[1] + 0.5 * gt_w
# Anchor
a_h = a[2] - a[0]
a_w = a[3] - a[1]
a_center_y = a[0] + 0.5 * a_h
a_center_x = a[1] + 0.5 * a_w
# Compute the bbox refinement that the RPN should predict.
rpn_bbox[ix] = [
(gt_center_y - a_center_y) / a_h,
(gt_center_x - a_center_x) / a_w,
np.log(gt_h / a_h),
np.log(gt_w / a_w),
]
# Normalize
rpn_bbox[ix] /= config.RPN_BBOX_STD_DEV
ix += 1
return rpn_match, rpn_bbox
def build_rpn_targets(anchors, gt_class_ids, gt_boxes, config):
"""Given the anchors and GT boxes, compute overlaps and identify positive
anchors and deltas to refine them to match their corresponding GT boxes.
anchors: [num_anchors, (y1, x1, y2, x2)]
gt_class_ids: [num_gt_boxes] Integer class IDs.
gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)]
Returns:
rpn_match: [N] (int32) matches between anchors and GT boxes.
1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas.
"""
# RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral
rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32)
# RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))]
rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4))
crowd_ix = np.where(gt_class_ids < 0)[0]
if crowd_ix.shape[0] > 0:
# Filter out crowds from ground truth class IDs and boxes
non_crowd_ix = np.where(gt_class_ids > 0)[0]
crowd_boxes = gt_boxes[crowd_ix]
gt_class_ids = gt_class_ids[non_crowd_ix]
gt_boxes = gt_boxes[non_crowd_ix]
# Compute overlaps with crowd boxes [anchors, crowds]
crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes)
crowd_iou_max = np.amax(crowd_overlaps, axis=1)
no_crowd_bool = (crowd_iou_max < 0.001)
else:
# All anchors don't intersect a crowd
no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool)
# Compute overlaps [num_anchors, num_gt_boxes]
overlaps = utils.compute_overlaps(anchors, gt_boxes)
anchor_iou_argmax = np.argmax(overlaps, axis=1)
anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
rpn_match[(anchor_iou_max < 0.3) & (no_crowd_bool)] = -1
# gt_box 匹配到anchors 和anchor 和gt_box>0.7都记为正例
gt_iou_argmax = np.argmax(overlaps, axis=0)
rpn_match[gt_iou_argmax] = 1
rpn_match[anchor_iou_max >= 0.7] = 1
# 防止正例超过要训练anchor的一半
ids = np.where(rpn_match == 1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2)
if extra > 0:
# Reset the extra ones to neutral
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
# 防止负例和正例之和超过总数
ids = np.where(rpn_match == -1)[0]
extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE -
np.sum(rpn_match == 1))
if extra > 0:
ids = np.random.choice(ids, extra, replace=False)
rpn_match[ids] = 0
ids = np.where(rpn_match == 1)[0]
ix = 0 # index into rpn_bbox
# TODO: use box_refinement() rather than duplicating the code here
for i, a in zip(ids, anchors[ids]):
# Closest gt box (it might have IoU < 0.7)
gt = gt_boxes[anchor_iou_argmax[i]]
# Convert coordinates to center plus width/height.
# GT Box
gt_w = gt[2] - gt[0]
gt_h = gt[3] - gt[1]
gt_center_x = gt[0] + 0.5 * gt_h
gt_center_y = gt[1] + 0.5 * gt_w
# Anchor
a_w = a[2] - a[0]
a_h = a[3] - a[1]
a_center_x = a[0] + 0.5 * a_h
a_center_y = a[1] + 0.5 * a_w
# Compute the bbox refinement that the RPN should predict.
rpn_bbox[ix] = [
(gt_center_x - a_center_x) / a_w,
(gt_center_y - a_center_y) / a_h,
np.log(gt_w / a_w),
np.log(gt_h / a_h),
]
# Normalize
rpn_bbox[ix] /= config.RPN_BBOX_STD_DEV
ix += 1
return rpn_match, rpn_bbox
