def create_frames_with_overlays(self, video, mode='gt'):
"""Build list of individual frames in a video with masks overlayed."""
# Overlay masks on top of images
frames_with_overlays = []
num_frames = self.videos[video]
alpha = 0.6
if mode == 'gt' or mode == 'gt_warped':
masks = self.masks_train
overlay_color = (255, 0, 0) # red
bbox_color = (0, 255, 255) # cyan
else:
masks = self.warped_prev_masks_train
overlay_color = (0, 255, 0) # green
bbox_color = (255, 255, 0) # yellow
for frame_number in range(num_frames):
frame_idx = self.video_frame_idx[video + '_' + str(frame_number)]
frame_with_overlay = visualize.apply_mask(self.images_train[frame_idx], masks[frame_idx], overlay_color, alpha)
bbox = bboxes.extract_bbox(masks[frame_idx])
frame_with_overlay = visualize.draw_box(frame_with_overlay, bbox, bbox_color) # y1, x1, y2, x2 order
frames_with_overlays.append(frame_with_overlay)
if mode == 'gt_warped':
masks = self.warped_prev_masks_train
overlay_color = (0, 255, 0) # green
bbox_color = (255, 255, 0) # yellow
for frame_number in range(num_frames):
frame_idx = self.video_frame_idx[video + '_' + str(frame_number)]
frame_with_overlay = visualize.apply_mask(frames_with_overlays[frame_number], masks[frame_idx], overlay_color, alpha, in_place = True)
bbox = bboxes.extract_bbox(masks[frame_idx])
frame_with_overlay = visualize.draw_box(frame_with_overlay, bbox, bbox_color) # y1, x1, y2, x2 order
frames_with_overlays[frame_number] = frame_with_overlay
return frames_with_overlays
rows = np.ascontiguousarray(rois[b]).view(
np.dtype((np.void, rois.dtype.itemsize * rois.shape[-1])))
_, idx = np.unique(rows, return_index=True)
print("Unique ROIs: {} out of {}".format(len(idx), rois.shape[1]))
# In[20]:
if random_rois:
# Dispalay ROIs and corresponding masks and bounding boxes
ids = random.sample(range(rois.shape[1]), 8)
images = []
titles = []
for i in ids:
image = visualize.draw_box(sample_image.copy(),
rois[b,
i, :4].astype(np.int32), [255, 0, 0])
image = visualize.draw_box(image, refined_rois[i].astype(np.int64),
[0, 255, 0])
images.append(image)
titles.append("ROI {}".format(i))
images.append(mask_specific[i] * 255)
titles.append(dataset.class_names[mrcnn_class_ids[b, i]][:20])
display_images(images, titles, cols=4, cmap="Blues", interpolation="none")
# In[21]:
# Check ratio of positive ROIs in a set of images.
if random_rois:
limit = 10
def bounding_boxes(t, t2, sm_buf, med_buf, lar_buf):
"""
Bounding box generator for ground truth segmented lesions 't'. Heavily modified from Tanya's
lesion_counter.
:param t: ground truth labels, shape=(NxMxO)
:type t: int16
:params sm_buf, med_buf, lar_buf: buffers to be applied for respective lesion sizes (voxel padding)
:return: dict
"""
# Remove small lesions and init bounding boxes and lesion classes
t = remove_tiny_les(t, nvox=2)
labels = {}
nles = {}
labels, nles = ndimage.label(t)
print('Number of lesions', nles)
boxes = np.zeros([nles + 1, 6], dtype=np.int32)
classes = np.zeros([nles + 1, 1], dtype=np.int32)
nb_lesions = nles
# Look for all the voxels associated with a particular lesion, then bound on x, y, z axis
for i in range(1, nles + 1):
t[labels != i] = 0
t[labels == i] = 1
# Now we classify the lesion and apply a buffer based on the lesion class (CHANGE LATER??)
lesion_size = np.sum(t[labels == i])
classes[i, 0] = get_lesion_bin(lesion_size)
x_indicies = np.where(np.any(t, axis=0))[0]
y_indicies = np.where(np.any(t, axis=1))[0]
z_indicies = []
for lesion_slice in range(t.shape[-1]):
if np.any(t[..., lesion_slice]):
z_indicies.append(lesion_slice)
z_indicies = np.asarray(z_indicies)
if x_indicies.shape[0]:
x1, x2 = x_indicies[[0, -1]]
y1, y2 = y_indicies[[0, -1]]
z1, z2 = z_indicies[[0, -1]]
x2 += 1
y2 += 1
z2 += 1
if classes[i] == 0:
x1 -= sm_buf
x2 += sm_buf
y1 -= sm_buf
y2 += sm_buf
z1 -= sm_buf
z2 += sm_buf
elif classes[i] == 1:
x1 -= med_buf
x2 += med_buf
y1 -= med_buf
y2 += med_buf
z1 -= med_buf
z2 += med_buf
else:
x1 -= lar_buf
x2 += lar_buf
y1 -= lar_buf
y2 += lar_buf
z1 -= lar_buf
z2 += lar_buf
else:
# No mask for this instance
x1, x2, y1, y2, z1, z2 = 0, 0, 0, 0, 0, 0
boxes[i] = np.array([y1, x1, y2, x2, z1, z2])
# Reset ground truth mask and then we can draw boxes
for i in range(1, nb_lesions + 1):
t[labels == i] = 1
for brain_slice in range(18, 50):
for les in range(1, nles + 1):
if brain_slice in range(boxes[les, 4], boxes[les, 5]):
#img2 = visualize.draw_boxes(t2[:,:,35], boxes=boxes[les, 0:4], masks=t[:,:,35])
if classes[les] == 0:
img = visualize.draw_box(t[:, :, brain_slice],
boxes[les, 0:4],
color=0.2)
elif classes[les] == 1:
img = visualize.draw_box(t[:, :, brain_slice],
boxes[les, 0:4],
color=0.4)
else:
img = visualize.draw_box(t[:, :, brain_slice],
boxes[les, 0:4],
color=0.7)
else:
img = t[:, :, brain_slice]
imgplt = plt.imshow(img)
print(brain_slice)
plt.show()
plt.savefig('slice.pdf')
return {'classes': classes, 'boxes': boxes}