def prepare_image(self, image_id):
"""use config to processing coco image size and others,
augment: (deprecated. Use augmentation instead). If true, apply random
image augmentation. Currently, only horizontal flipping is offered.
augmentation: Optional. An imgaug (https://github.com/aleju/imgaug) augmentation.
For example, passing imgaug.augmenters.Fliplr(0.5) flips images
right/left 50% of the time.
Returns:
image: [height, width, 3]
image_meta: the original shape of the image and resizing and cropping.
class_ids: [instance_count] Integer class IDs
bbox: [instance_count, (y1, x1, y2, x2)]
mask: [height, width, instance_count]. The height and width are those
of the image.
gt_y: [instance_count]
gt_x: [instance_count]
vector_mask: [height, width, 2*class_num]. Set pixel relative center vector.
"""
# Load image and mask
image = self.load_image(image_id=image_id)
mask, class_ids = self.load_mask(image_id=image_id)
original_shape = image.shape
# print(original_shape)
# print(type(original_shape))
image, window, scale, padding, crop = cocoutils.resize_image(
image,
min_dim=self.config.IMAGE_MIN_DIM,
min_scale=self.config.IMAGE_MIN_SCALE,
max_dim=self.config.IMAGE_MAX_DIM,
mode=self.config.IMAGE_RESIZE_MODE)
mask = cocoutils.resize_mask(mask, scale, padding, 0, crop)
_idx = np.sum(mask, axis=(0, 1)) > 16
class_ids = class_ids[_idx]
if len(class_ids) != 0:
# print(class_ids)
# [y, x, num_instance]
mask = mask[:, :, _idx]
# print(np.amax(mask, axis=(0, 1)))
# Bounding boxes. Note that some boxes might be all zeros
# if the corresponding mask got cropped out.
# bbox: [num_instances, (y1, x1, y2, x2)]
bbox = cocoutils.extract_bboxes(mask)
gt_cy, gt_cx = cocoutils.gravity_center(mask)
return image, class_ids, bbox, mask, gt_cy, gt_cx
print("return nothing")
return None
def generate_stuff(config, semantic_segmentation):
# [batch, num_detections, (y1, x1, y2, x2)]
# [batch, num_detections, h, w]
if config.GPU_COUNT:
segments = semantic_segmentation.squeeze(0).data.cpu().numpy()
else:
segments = semantic_segmentation.squeeze(0).data.numpy()
pred = np.argmax(segments, axis=0)
results = []
masks = []
for i in range(config.THING_NUM_CLASSES,
config.THING_NUM_CLASSES + config.STUFF_NUM_CLASSES):
mask = (pred == i)
mask = np.where(mask, 1, 0)
count = np.sum(mask)
# if count>5000:
if count > config.STUFF_THRESHOLD:
mask = mask.reshape(1, mask.shape[0], mask.shape[1])
reshaped_mask = mask.transpose(1, 2, 0)
bbox = extract_bboxes(reshaped_mask) # 1,4
r = np.array([[i]], dtype=np.float32) # 1,1
r = np.concatenate((bbox, r), axis=1) # 1,5
results.append(r)
# plt.figure()
# plt.imshow(mask[0])
# plt.show()
masks.append(mask) # (1,500,500)
if len(results) > 0 and len(masks) > 0:
results = Variable(torch.from_numpy(np.concatenate(results)).float())
masks = Variable(torch.from_numpy(np.concatenate(masks)))
else:
results = Variable(torch.from_numpy(np.array(results)).float())
masks = Variable(torch.from_numpy(np.array(masks)).float())
return pred, results, masks
def prepare_image(self, image_id, augment=False, augmentation=None):
"""use config to processing coco image size and others,
augment: (deprecated. Use augmentation instead). If true, apply random
image augmentation. Currently, only horizontal flipping is offered.
augmentation: Optional. An imgaug (https://github.com/aleju/imgaug) augmentation.
For example, passing imgaug.augmenters.Fliplr(0.5) flips images
right/left 50% of the time.
Returns:
image: [height, width, 3]
image_meta: the original shape of the image and resizing and cropping.
class_ids: [instance_count] Integer class IDs
bbox: [instance_count, (y1, x1, y2, x2)]
mask: [height, width, instance_count]. The height and width are those
of the image.
gt_y: [instance_count]
gt_x: [instance_count]
vector_mask: [height, width, 2*class_num]. Set pixel relative center vector.
"""
# Load image and mask
image = self.load_image(image_id)
mask, class_ids = self.load_mask(image_id)
# original_shape = image.shape
# print(original_shape)
# print(type(original_shape))
image, window, scale, padding, crop = cocoutils.resize_image(
image,
min_dim=self.config.IMAGE_MIN_DIM,
min_scale=self.config.IMAGE_MIN_SCALE,
max_dim=self.config.IMAGE_MAX_DIM,
mode=self.config.IMAGE_RESIZE_MODE)
mask = cocoutils.resize_mask(mask, scale, padding, 0, crop)
# Random horizontal flips.
# TODO: will be removed in a future update in favor of augmentation
if self.augment:
logging.warning(
"'augment' is deprecated. Use 'augmentation' instead.")
if random.randint(0, 1):
image = np.fliplr(image)
mask = np.fliplr(mask)
# Augmentation
# This requires the imgaug lib (https://github.com/aleju/imgaug)
if self.augmentation:
import imgaug
# Augmenters that are safe to apply to masks
# Some, such as Affine, have settings that make them unsafe, so always
# test your augmentation on masks
MASK_AUGMENTERS = [
"Sequential", "SomeOf", "OneOf", "Sometimes", "Fliplr",
"Flipud", "CropAndPad", "Affine", "PiecewiseAffine"
]
def hook(images, augmenter, parents, default):
"""Determines which augmenters to apply to masks."""
return augmenter.__class__.__name__ in MASK_AUGMENTERS
# Store shapes before augmentation to compare
image_shape = image.shape
mask_shape = mask.shape
# Make augmenters deterministic to apply similarly to images and masks
det = augmentation.to_deterministic()
image = det.augment_image(image)
# Change mask to np.uint8 because imgaug doesn't support np.bool
mask = det.augment_image(mask.astype(np.uint8),
hooks=imgaug.HooksImages(activator=hook))
# Verify that shapes didn't change
assert image.shape == image_shape, "Augmentation shouldn't change image size"
assert mask.shape == mask_shape, "Augmentation shouldn't change mask size"
# Change mask back to bool
mask = mask.astype(np.bool)
_idx = np.sum(mask, axis=(0, 1)) > 48
# print(_idx)
class_ids = class_ids[_idx]
if len(class_ids) != 0:
# print(class_ids)
# [y, x, num_instance]
mask = mask[:, :, _idx]
# print(np.amax(mask, axis=(0, 1)))
# Bounding boxes. Note that some boxes might be all zeros
# if the corresponding mask got cropped out.
# bbox: [num_instances, (y1, x1, y2, x2)]
bbox = cocoutils.extract_bboxes(mask)
gt_cy, gt_cx = cocoutils.gravity_center(mask)
# Image meta data
# image_meta = cocoutils.compose_image_meta(image_id, original_shape, image.shape, window, scale)
# vector_mask = self.vector_mask(self.num_classes, class_ids, mask, gt_cx, gt_cy, image.shape, bbox)
return image, class_ids, bbox, gt_cy, gt_cx, mask
return None
def load_image_gt(dataset,
config,
image_id,
augment=False,
augmentation=None,
use_mini_mask=False):
# 载入图片和语义分割效果
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
# print("\nbefore:",image_id,np.shape(mask),np.shape(class_ids))
# 原始shape
original_shape = image.shape
# 获得新图片，原图片在新图片中的位置，变化的尺度，填充的情况等
image, window, scale, padding, crop = utils.resize_image(
image,
min_dim=config.IMAGE_MIN_DIM,
min_scale=config.IMAGE_MIN_SCALE,
max_dim=config.IMAGE_MAX_DIM,
mode=config.IMAGE_RESIZE_MODE)
mask = utils.resize_mask(mask, scale, padding, crop)
# print("\nafter:",np.shape(mask),np.shape(class_ids))
# print(np.shape(image),np.shape(mask))
# 可以把图片进行翻转
if augment:
logging.warning("'augment' is deprecated. Use 'augmentation' instead.")
if random.randint(0, 1):
image = np.fliplr(image)
mask = np.fliplr(mask)
if augmentation:
import imgaug
# 可用于图像增强
MASK_AUGMENTERS = [
"Sequential", "SomeOf", "OneOf", "Sometimes", "Fliplr", "Flipud",
"CropAndPad", "Affine", "PiecewiseAffine"
]
def hook(images, augmenter, parents, default):
"""Determines which augmenters to apply to masks."""
return augmenter.__class__.__name__ in MASK_AUGMENTERS
image_shape = image.shape
mask_shape = mask.shape
det = augmentation.to_deterministic()
image = det.augment_image(image)
mask = det.augment_image(mask.astype(np.uint8),
hooks=imgaug.HooksImages(activator=hook))
assert image.shape == image_shape, "Augmentation shouldn't change image size"
assert mask.shape == mask_shape, "Augmentation shouldn't change mask size"
mask = mask.astype(np.bool)
# 检漏，防止某些层内部实际上不存在语义分割情况
_idx = np.sum(mask, axis=(0, 1)) > 0
# print("\nafterer:",np.shape(mask),np.shape(_idx))
mask = mask[:, :, _idx]
class_ids = class_ids[_idx]
# 找到mask对应的box
bbox = utils.extract_bboxes(mask)
active_class_ids = np.zeros([dataset.num_classes], dtype=np.int32)
source_class_ids = dataset.source_class_ids[dataset.image_info[image_id]
["source"]]
active_class_ids[source_class_ids] = 1
if use_mini_mask:
mask = utils.minimize_mask(bbox, mask, config.MINI_MASK_SHAPE)
# 生成Image_meta
image_meta = utils.compose_image_meta(image_id, original_shape,
image.shape, window, scale,
active_class_ids)
return image, image_meta, class_ids, bbox, mask
def parse_fn(image_id,
dataset,
anchors_path,
augmentation=None,
dtype=np.float32,
max_num_boxes_per_image=20,
image_size=416):
"""Load and return ground truth data for an image (image, mask, bounding boxes)."""
image = dataset.load_image(image_id)
# original_shape = image.shape
image, window, scale, padding, crop = utils.resize_image(
image, min_dim=0, min_scale=0, max_dim=image_size, mode='square')
mask, class_ids = dataset.load_mask(image_id)
mask = utils.resize_mask(mask, scale, padding, crop)
if augmentation:
import imgaug
# Augmenters that are safe to apply to masks
# Some, such as Affine, have settings that make them unsafe, so always
# test your augmentation on masks
MASK_AUGMENTERS = [
"Sequential", "SomeOf", "OneOf", "Sometimes", "Fliplr", "Flipud",
"CropAndPad", "Affine", "PiecewiseAffine"
]
def hook(images, augmenter, parents, default):
"""Determines which augmenters to apply to masks."""
return augmenter.__class__.__name__ in MASK_AUGMENTERS
# Store shapes before augmentation to compare
image_shape = image.shape
mask_shape = mask.shape
# Make augmenters deterministic to apply similarly to images and masks
det = augmentation.to_deterministic()
image = det.augment_image(image)
# Change mask to np.uint8 because imgaug doesn't support np.bool
mask = det.augment_image(mask.astype(np.uint8),
hooks=imgaug.HooksImages(activator=hook))
# Verify that shapes didn't change
assert image.shape == image_shape, "Augmentation shouldn't change image size"
assert mask.shape == mask_shape, "Augmentation shouldn't change mask size"
# Change mask back to bool
mask = mask.astype(np.bool)
# Note that some boxes might be all zeros if the corresponding mask got cropped out.
# and here is to filter them out
_idx = np.sum(mask, axis=(0, 1)) > 0
mask = mask[:, :, _idx]
class_ids = class_ids[_idx]
# Bounding boxes. Note that some boxes might be all zeros
# if the corresponding mask got cropped out.
# bbox: [num_instances, (y1, x1, y2, x2)]
bbox = utils.extract_bboxes(mask)
if mask.shape[-1] > max_num_boxes_per_image:
ids = np.random.choice(np.arange(mask.shape[-1]),
max_num_boxes_per_image,
replace=False)
class_ids = class_ids[ids]
bbox = bbox[ids, :]
# confs = np.ones((bbox.shape[0], 1), dtype=dtype)
# bbox = np.concatenate([bbox, confs], axis=-1)
# Active classes
# Different datasets have different classes, so track the
# classes supported in the dataset of this image.
# active_class_ids = np.zeros([dataset.num_classes], dtype=np.int32)
# source_class_ids = dataset.source_class_ids[dataset.image_info[image_id]["source"]]
# active_class_ids[source_class_ids] = 1
# image_meta = utils.compose_image_meta(image_id, original_shape, image.shape,
# window, scale, active_class_ids)
# image_meta.astype(dtype)
# gt_mask = np.zeros((mask.shape[0], mask.shape[1], 20), mask.dtype)
gt_class_ids = np.zeros(max_num_boxes_per_image, class_ids.dtype)
gt_bbox = np.zeros((max_num_boxes_per_image, bbox.shape[1]), bbox.dtype)
# gt_data = np.zeros((max_num_boxes_per_image, bbox.shape[1] + dataset.num_classes), dtype=dtype)
if class_ids.shape[0] > 0:
gt_class_ids[:class_ids.shape[0]] = class_ids
# gt_mask[:, :, :mask.shape[-1]] = mask
gt_bbox[:bbox.shape[0], :] = bbox
gt_class_ids = np.expand_dims(gt_class_ids, axis=-1).astype(dtype)
gt_bbox = np.concatenate([gt_bbox, gt_class_ids], axis=-1)
anchors = utils.get_anchors(anchors_path)
anchors = np.array(anchors, dtype=np.float32)
boxes_yx = (gt_bbox[:, 0:2] + gt_bbox[:, 2:4]) // 2
boxes_hw = gt_bbox[:, 2:4] - gt_bbox[:, 0:2]
gt_bbox[:, 0] = boxes_yx[..., 1] / image_size
gt_bbox[:, 1] = boxes_yx[..., 0] / image_size
gt_bbox[:, 2] = boxes_hw[..., 1] / image_size
gt_bbox[:, 3] = boxes_hw[..., 0] / image_size
hw = np.expand_dims(boxes_hw, -2)
anchors_broad = np.expand_dims(anchors, 0)
anchor_maxes = anchors_broad / 2.
anchor_mins = -anchor_maxes
box_maxes = hw / 2.
box_mins = -box_maxes
intersect_mins = np.maximum(box_mins, anchor_mins)
intersect_maxes = np.minimum(box_maxes, anchor_maxes)
intersect_hw = np.maximum(intersect_maxes - intersect_mins, 0.)
intersect_area = intersect_hw[..., 0] * intersect_hw[..., 1]
box_area = hw[..., 0] * hw[..., 1]
anchor_area = anchors[..., 0] * anchors[..., 1]
iou = intersect_area / (box_area + anchor_area - intersect_area)
best_anchors = np.argmax(iou, axis=-1)
# TODO: write a function to calculate the stride automatically.
large_obj_image_size = image_size // 32
medium_obj_image_size = image_size // 16
small_obj_image_size = image_size // 8
large_obj_detectors, large_obj_boxes = get_detector_heatmap_each_scale(
gt_bbox,
best_anchors_=best_anchors,
anchors_mask=[6, 7, 8],
grid_size=(large_obj_image_size, large_obj_image_size),
num_classes=dataset.num_classes)
medium_obj_detectors, medium_obj_boxes = get_detector_heatmap_each_scale(
gt_bbox,
best_anchors_=best_anchors,
anchors_mask=[3, 4, 5],
grid_size=(medium_obj_image_size, medium_obj_image_size),
num_classes=dataset.num_classes)
small_obj_detectors, small_obj_boxes = get_detector_heatmap_each_scale(
gt_bbox,
best_anchors_=best_anchors,
anchors_mask=[0, 1, 2],
grid_size=(small_obj_image_size, small_obj_image_size),
num_classes=dataset.num_classes)
yolo_true_data = np.concatenate(
[large_obj_detectors, medium_obj_detectors, small_obj_detectors],
axis=0).reshape([-1])
yolo_true_boxes = np.concatenate(
[large_obj_boxes, medium_obj_boxes, small_obj_boxes],
axis=0).reshape([-1])
yolo_gt = np.concatenate([yolo_true_data, yolo_true_boxes], axis=-1)
return image.astype(dtype) / 255., yolo_gt.astype(dtype)
def load_image_gt(dataset,
config,
image_id,
augment=False,
use_mini_mask=False):
"""Load and return ground truth data for an image (image, mask, bounding boxes).
augment: If true, apply random image augmentation. Currently, only
horizontal flipping is offered.
use_mini_mask: If False, returns full-size masks that are the same height
and width as the original image. These can be big, for example
1024x1024x100 (for 100 instances). Mini masks are smaller, typically,
224x224 and are generated by extracting the bounding box of the
object and resizing it to MINI_MASK_SHAPE.
Returns:
image: [height, width, 3]
shape: the original shape of the image before resizing and cropping.
class_ids: [instance_count] Integer class IDs
bbox: [instance_count, (y1, x1, y2, x2)]
mask: [height, width, instance_count]. The height and width are those
of the image unless use_mini_mask is True, in which case they are
defined in MINI_MASK_SHAPE.
"""
# Load image and mask
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
shape = image.shape
image, window, scale, padding = utils.resize_image(
image,
min_dim=config.TRAIN.IMAGE_MIN_DIM,
max_dim=config.TRAIN.IMAGE_MAX_DIM,
padding=config.TRAIN.IMAGE_PADDING)
mask = utils.resize_mask(mask, scale, padding)
# Random horizontal flips.
if augment:
if random.randint(0, 1):
image = np.fliplr(image)
mask = np.fliplr(mask)
# Bounding boxes. Note that some boxes might be all zeros
# if the corresponding mask got cropped out.
# bbox: [num_instances, (y1, x1, y2, x2)]
bbox = utils.extract_bboxes(mask)
# Active classes
# Different datasets have different classes, so track the
# classes supported in the dataset of this image.
active_class_ids = np.zeros([dataset.num_classes], dtype=np.int32)
source_class_ids = dataset.source_class_ids[dataset.image_info[image_id]
["source"]]
active_class_ids[source_class_ids] = 1
# Resize masks to smaller size to reduce memory usage
if use_mini_mask:
mask = utils.minimize_mask(bbox, mask, config.MRCNN.MINI_MASK_SHAPE)
# Image meta data
image_meta = compose_image_meta(image_id, shape, window, active_class_ids)
return image, image_meta, class_ids, bbox, mask
def load_image_gt(dataset, config, image_id, augment=False, use_mini_mask=False):
# Load image and mask
image_name = dataset.image_info[str(image_id)]['image_name']
# print(image_name)
image = dataset.load_image(image_id)
shape = image.shape
image, window, scale, padding = resize_image(
image,
min_dim=config.IMAGE_MIN_DIM,
max_dim=config.IMAGE_MAX_DIM,
padding=config.IMAGE_PADDING)
image_meta = compose_image_meta(image_id, shape, window)
thing_mask, thing_class_ids, stuff_mask, stuff_class_ids, influence_mask, influence_class_ids = dataset.load_mask(image_id)
thing_mask = resize_mask(thing_mask, scale, padding) # 1024
stuff_mask = resize_mask(stuff_mask, scale, padding) # 1024
influence_mask = resize_mask(influence_mask, scale, padding) # 1024
influence_mask = resize_map(influence_mask, 1 / 8) # 1024 -> 128
# Resize masks to smaller size to reduce memory usage
thing_bbox = extract_bboxes(thing_mask)
stuff_bbox = extract_bboxes(stuff_mask)
influence_bbox = extract_bboxes(influence_mask)
if use_mini_mask:
thing_mask = minimize_mask(
thing_bbox, thing_mask, config.MINI_MASK_SHAPE)
stuff_mask = minimize_mask(
stuff_bbox, stuff_mask, config.MINI_MASK_SHAPE)
segmentation = skimage.io.imread(os.path.join(dataset.annotation_dir, image_name.replace("jpg", "png")))
semantic_label = np.zeros_like(segmentation)
segmentation_instance_id_map=rgb2id(segmentation)
instance_id_list=list(dataset.image_info[str(image_id)]['instances'].keys())
for instance_id in instance_id_list:
instance=dataset.image_info[str(image_id)]['instances'][instance_id]
instance_mask=segmentation_instance_id_map==int(instance_id)
semantic_label[instance_mask]=dataset.category_info[str(instance['category_id'])]['class_id']
semantic_label=semantic_label[:,:,0]
semantic_label_h = semantic_label.shape[0]
semantic_label_w = semantic_label.shape[1]
semantic_label_scale = min(500 / semantic_label_h, 500 / semantic_label_w)
semantic_label = scipy.misc.imresize(semantic_label, (round(semantic_label_h * semantic_label_scale), round(semantic_label_w * semantic_label_scale)), interp="nearest")
h, w = semantic_label.shape[:2]
top_pad = (500 - h) // 2
bottom_pad = 500 - h - top_pad
left_pad = (500 - w) // 2
right_pad = 500 - w - left_pad
padding = [(top_pad, bottom_pad), (left_pad, right_pad)]
semantic_label = np.pad(semantic_label, padding, mode='constant', constant_values=0)
image_info = dataset.image_info[str(image_id)]
# Random horizontal flips.
if augment:
if random.randint(0, 1):
image = np.fliplr(image)
thing_mask = np.fliplr(thing_mask)
semantic_label = np.fliplr(semantic_label)
segmentation = np.fliplr(segmentation)
return image, image_meta, thing_class_ids, thing_bbox, thing_mask, stuff_class_ids, stuff_bbox, stuff_mask, \
semantic_label, segmentation, image_info, influence_class_ids, influence_bbox, influence_mask
image_ids = dataset.image_ids
print(len(image_ids))
for idx, image_id in enumerate(image_ids):
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
visualize.display_top_masks(image,
mask,
class_ids,
dataset.class_names,
img_idx=idx)
# Load random image and mask.
for idx, image_id in enumerate(image_ids):
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id ", image_id, dataset.image_reference(image_id))
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image,
bbox,
mask,
class_ids,
dataset.class_names,
img_idx=idx)
# Generate Anchors