def CleanMask_v1(mask):
# remove small objects and fill holes
mask = (mask > .5).astype(np.int)
mask = binary_fill_holes(mask)
lbl_mask, numObj = scipy_label(mask)
processed_mask = np.zeros_like(mask)
minimum_cc_sum = .005 * np.prod(mask.shape)
for label in range(1, numObj + 1):
if np.sum(lbl_mask == label) > minimum_cc_sum:
processed_mask[lbl_mask == label] = 1
return processed_mask.astype(np.int)
def generate_submission(self):
test_datapath = '/data/Kaggle/test-png'
# Get list of files
img_files = natsorted(glob(join(test_datapath, '*.png')))
# load files into array
test_imgs = np.stack(
[self.LoadImgForTest(f, self.dims) for f in img_files])[...,
np.newaxis]
# Get predicted masks
tqdm.write('Getting mask predictions...')
masks = self.model.predict(test_imgs,
batch_size=self.batch_size,
verbose=1)
# data to write to csv
submission_data = []
# process mask
for ind, cur_file in enumerate(tqdm(img_files)):
cur_mask = masks[ind, ..., 0]
cur_im = test_imgs[ind, ..., 0]
cur_mask = (cur_mask > .5).astype(np.int)
cur_id = splitfile(cur_file)
processed_mask = CleanMask_v1(cur_mask)
lbl_mask, numObj = scipy_label(processed_mask)
if numObj > 0:
for label in range(1, numObj + 1):
temp_mask = np.zeros_like(cur_mask)
temp_mask[lbl_mask == label] = 1
temp_mask = cv2.resize(temp_mask.astype(np.float),
(1024, 1024))
temp_mask[temp_mask < .5] = 0
temp_mask[temp_mask > 0] = 255
temp_mask = np.transpose(temp_mask)
cur_rle = mask2rle(temp_mask, 1024, 1024)
submission_data.append([cur_id, cur_rle])
else:
cur_rle = -1
submission_data.append([cur_id, cur_rle])
# write to csv
# generate time-stamped filename
timestamp = datetime.datetime.now().strftime("%Y_%m_%d_%H%M")
csv_filename = 'TestSubmission_{}'.format(timestamp)
tqdm.write('Writing csv...')
with open(csv_filename, mode='w') as f:
writer = csv.writer(f, delimiter=',')
for row in tqdm(submission_data):
writer.writerow(row)
print('Done')
def GetSubData(file, label, mask):
mask = mask[..., 0]
mask = (mask > .5).astype(np.int)
fid = splitfile(file)
if label == 0:
return [fid, -1]
processed_mask = CleanMask_v1(mask)
lbl_mask, numObj = scipy_label(processed_mask)
if numObj > 0:
processed_mask[processed_mask > 0] = 255
processed_mask = np.transpose(processed_mask)
rle = mask2rle(processed_mask, 1024, 1024)
else:
rle = -1
return [fid, rle]
def pipeline(img):
img_draw_search = img.copy()
img_draw_cars = img.copy()
img_processed = process_image(img)
img_search = img_processed[p_search.ystart:p_search.ystop, :, :]
shape = img_search.shape
img_search = cv2.resize(
img_search,
(np.int(shape[1] / p_search.scale), np.int(shape[0] / p_search.scale)))
hog_features = get_hog_features(img_search, p_features)[0]
heatmap = slide_and_search(img_search, img_draw_search, hog_features,
classifier, p_search, p_features)
heatmaps.update(heatmap)
heatmap_sum = heatmaps.get_sum()
heatmap_thresh = heatmap_sum.copy()
apply_threshold(heatmap_thresh, 25)
boxes = scipy_label(heatmap_thresh)
draw_car_boxes(img_draw_cars, boxes)
return img_draw_search, img_draw_cars, heatmap_sum
def detect_largest_umap_areas_slice(slice_u_map, structure):
binary_map = np.zeros(slice_u_map.shape).astype(np.bool)
mask = slice_u_map != 0
binary_map[mask] = True
binary_structure = generate_binary_structure(binary_map.ndim, connectivity=2)
bin_labels, num_of_objects = scipy_label(binary_map, binary_structure)
blob_area_sizes = []
blob_after_erosion_sizes = []
if num_of_objects >= 1:
for i in np.arange(1, num_of_objects + 1):
binary_map = np.zeros(bin_labels.shape).astype(np.bool)
binary_map[bin_labels == i] = 1
blob_area_sizes.append(np.count_nonzero(binary_map))
remaining_blob = binary_erosion(binary_map, structure)
blob_after_erosion_sizes.append(np.count_nonzero(remaining_blob))
blob_area_sizes = np.array(blob_area_sizes)
blob_area_sizes[::-1].sort()
# print(blob_after_erosion_sizes)
blob_after_erosion_sizes = np.array(blob_after_erosion_sizes)
blob_after_erosion_sizes[::-1].sort()
return blob_area_sizes, blob_after_erosion_sizes
def _is_contiguous(self, arr):
_, num_features = scipy_label(arr)
return num_features == 1
def getLargestCC(segmentation):
labels, count = scipy_label(segmentation, structure=np.ones((3, 3, 3)))
assert( labels.max() != 0 ) # assume at least 1 CC
largestCC = labels == np.argmax(np.bincount(labels.flat)[1:])+ 1
return largestCC
# Get predicted masks
print('Getting predicted masks...')
masks = model.predict(test_imgs, batch_size=batch_size, verbose=1)
# data to write to csv
submission_data = []
# process mask
tqdm.write('Processing masks...')
for ind, cur_file in enumerate(tqdm(img_files)):
cur_mask = masks[ind, ..., 0]
cur_im = test_imgs[ind, ..., 0]
cur_mask = (cur_mask > .5).astype(np.int)
cur_id = splitfile(cur_file)
processed_mask = CleanMask_v1(cur_mask)
lbl_mask, numObj = scipy_label(processed_mask)
if numObj > 0:
for label in range(1, numObj + 1):
temp_mask = np.zeros_like(cur_mask)
temp_mask[lbl_mask == label] = 1
temp_mask = cv2.resize(temp_mask.astype(np.float), (1024, 1024))
temp_mask[temp_mask < .5] = 0
temp_mask[temp_mask > 0] = 255
temp_mask = np.transpose(temp_mask)
cur_rle = mask2rle(temp_mask, 1024, 1024)
submission_data.append([cur_id, cur_rle])
else:
cur_rle = -1
submission_data.append([cur_id, cur_rle])
# write to csv
def filter_candidates(img,
candidates,
cars,
heat_threshold=1,
max_smooth=10,
min_overlap=0.6,
min_area=1000):
"""
Given the list of candidates detected by sliding windows,
compute the heat map and extract the "canonical" car positions in the image
`cars` is a list of CarPosition objects
"""
heat = np.zeros_like(img[:, :, 0]).astype(np.float)
# add heat
for box in candidates:
heat[box[0][1]:box[1][1], box[0][0]:box[1][0]] += 1
# Apply threshold to help remove false positives
clipped_heat_map = np.copy(heat)
clipped_heat_map[clipped_heat_map <= heat_threshold] = 0
# Visualize the heatmap
heat_map = np.clip(clipped_heat_map, 0, 255)
# Find final boxes from heatmap using label function
labels = scipy_label(heat_map)
bboxes = get_labeled_bboxes(labels)
# miss by default
for c in cars:
c.misses += 1
# assign the bboxes to the cars
for bbox in bboxes:
if bbox_area(bbox) < min_area:
continue
if len(cars) == 0:
cars = [CarPosition(bbox, max_smooth=max_smooth)]
else:
overlaps = [c.overlap(bbox) for c in cars]
idx = np.argmax(overlaps)
if overlaps[idx] >= min_overlap:
cars[idx].update(bbox)
else:
cars.append(CarPosition(bbox, max_smooth=max_smooth))
# remove cars that miss a frame
max_misses = 3
min_count = 3
cars = [c for c in cars if c.misses < max_misses]
# merge near-by car positions
merged_cars = []
for i in range(len(cars)):
merged = True
for j in range(i + 1, len(cars)):
if cars[i].overlap(cars[j].position) >= min_overlap:
merged = False
break
if merged:
merged_cars.append(cars[i])
cars = merged_cars
# Draw the box on the image
car_img = np.copy(img)
for car in cars:
if car.count >= min_count:
pos = car.position
cv2.rectangle(car_img, pos[0], pos[1], (0, 0, 255), 6)
return cars, car_img, heat
def Process_video_frame(img):
global BoxHandler
global FrameCount
FrameCount = FrameCount + 1
#print("Frame Count :", FrameCount)
rectangle_boxes = []
#------------------------------------------------------
# find cars in 192 x 192 sliding windows (scale=3)
# searching area: (ystart - ystop) = 636 - 396 = 240 = 192 x 1.25 (Chosen)
#_, img_boxes, _, _ = \
#find_cars( img, conv_color='RGB2YCrCb', ystart=396, ystop=636, scale=3, svc=svc,
# X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
# spatial_size=spatial_size, hist_bins=hist_bins )
#
#rectangle_boxes.extend(img_boxes)
# find cars in 160 x 160 sliding windows (scale=2.5)
# searching area: (ystart - ystop) = 620 - 380 = 240 = 160 x 1.5 (Chosen)
# searching area: (ystart - ystop) = 630 - 430 = 200 = 160 x 1.25
_, img_boxes, _, _ = \
find_cars( img, conv_color='RGB2YCrCb', ystart=380, ystop=620, scale=2.5, svc=svc,
X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
spatial_size=spatial_size, hist_bins=hist_bins )
rectangle_boxes.extend(img_boxes)
# find cars in 128 x 128 sliding windows (scale=2)
# searching area: (ystart - ystop) = 636 - 380 = 256 = 128 x 2
# searching area: (ystart - ystop) = 572 - 380 = 192 = 128 x 1.5 (Chosen)
# searching area: (ystart - ystop) = 540 - 380 = 160 = 128 x 1.25
_, img_boxes, _, _ = \
find_cars( img, conv_color='RGB2YCrCb', ystart=380, ystop=540, scale=2, svc=svc,
X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
spatial_size=spatial_size, hist_bins=hist_bins )
rectangle_boxes.extend(img_boxes)
# find cars in 112 x 112 sliding windows (scale=1.75)
# searching area: (ystart - ystop) = 604 - 380 = 224 = 112 x 2 (Chosen)
# searching area: (ystart - ystop) = 508 - 368 = 140 = 112 x 1.25 (Chosen)
_, img_boxes, _, _ = \
find_cars( img, conv_color='RGB2YCrCb', ystart=368, ystop=508, scale=1.75, svc=svc,
X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
spatial_size=spatial_size, hist_bins=hist_bins )
rectangle_boxes.extend(img_boxes)
# find cars in 96 x 96 sliding windows (scale=1.5)
# searching area: (ystart - ystop) = 572 - 380 = 192 = 96 x 2
# searching area: (ystart - ystop) = 524 - 380 = 144 = 96 x 1.5
# searching area: (ystart - ystop) = 488 - 368 = 120 = 96 x 1.25 (Chosen)
# searching area: (ystart - ystop) = 464 - 368 = 96 = 96 x 1.0
_, img_boxes, _, _ = \
find_cars( img, conv_color='RGB2YCrCb', ystart=380, ystop=524, scale=1.5, svc=svc,
X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
spatial_size=spatial_size, hist_bins=hist_bins )
rectangle_boxes.extend(img_boxes)
# find cars in 64 x 64 sliding windows (scale=1)
# searching area: (ystart - ystop) = 508 - 380 = 128 = 64 x 2
# searching area: (ystart - ystop) = 496 - 400 = 96 = 64 x 1.5 (Chosen)
# searching area: (ystart - ystop) = 460 - 380 = 80 = 64 x 1.25
_, img_boxes, _, _ = \
find_cars( img, conv_color='RGB2YCrCb', ystart=400, ystop=496, scale=1, svc=svc,
X_scaler=X_scaler, orient=orient, pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
spatial_size=spatial_size, hist_bins=hist_bins )
rectangle_boxes.extend(img_boxes)
#-------------------------------------------------------
# Debug: Generate video with raw boxes drawn
if (boDebugMode == True) and (boOutputRawBoxes == True):
box_marked_img = draw_boxes(img, rectangle_boxes)
return box_marked_img
#------------------------------------------------------
BoxHandler.add_boxes(rectangle_boxes)
BoxHandler.add_heatmap(rectangle_boxes, img)
heatmap = np.zeros_like(img[:, :, 0]).astype(np.float)
heatmap = add_weight_to_heatmap(heatmap, BoxHandler.get_all_boxes())
heatmap = apply_threshold(heatmap, BoxHandler.get_bboxes_threshold())
heatmap_thresholds = BoxHandler.get_framecount_threshold()
#plt.imshow(heatmap_thresholds, cmap='hot')
#plt.show()
heatmap[(heatmap_thresholds == 0)] = 0
#-------------------------------------------------------
# Debug: Generate video with Heat Map + Raw Boxes
if boDebugMode == True:
heatmap_show = heatmap
zeroimage = np.zeros_like(img[:, :, 0]).astype(np.float)
heatmap_RGB = cv2.merge([heatmap_show * 3, zeroimage, zeroimage])
heatmap_RGB = heatmap_RGB.astype(np.uint8)
#plt.imshow(heatmap_RGB)
#plt.show()
debug_image = draw_boxes(img,
rectangle_boxes,
color=(0, 0, 255),
thick=2)
debug_image = cv2.addWeighted(debug_image, 1, heatmap_RGB, 0.8, 0)
labels = scipy_label(heatmap)
debug_image = draw_labeled_bboxes(debug_image,
labels,
color=(255, 0, 0),
thick=3)
return debug_image
#-------------------------------------------------------
else:
labels = scipy_label(heatmap)
refined_marked_img = draw_labeled_bboxes(img, labels)
labels = scipy_label(heatmap)
refined_marked_img = draw_labeled_bboxes(img, labels)
return refined_marked_img
images = [box_marked_img, heatmap]
titles = ["Marked Image", "Original Heat Map"]
#plt_fig = plt.figure(figsize=(10.8, 3), dpi=200)
#display_images_var1(plt_fig, 1, 2, images, titles)
f, ax = plt.subplots(1, 2, figsize=(10.4, 3.15))
f.tight_layout()
ax[0].imshow(images[0])
ax[0].set_title(titles[0], fontsize=14)
ax[1].imshow(images[1], cmap='hot')
ax[1].set_title(titles[1], fontsize=14)
plt.subplots_adjust(left=0.06, right=0.97, top=0.98, bottom=0.02)
#-------------------------------------------------------
heatmap = apply_threshold(heatmap, 2)
labels = scipy_label(heatmap)
refined_marked_img = draw_labeled_bboxes(img, labels)
images = [refined_marked_img, heatmap]
titles = ["Marked Image", "Thresholded Heat Map"]
#plt_fig = plt.figure(figsize=(10.8, 3), dpi=200)
#display_images_var1(plt_fig, 1, 2, images, titles)
f, ax = plt.subplots(1, 2, figsize=(10.4, 3.15))
f.tight_layout()
ax[0].imshow(images[0])
ax[0].set_title(titles[0], fontsize=14)
ax[1].imshow(images[1], cmap='hot')
ax[1].set_title(titles[1], fontsize=14)
plt.subplots_adjust(left=0.06, right=0.97, top=0.98, bottom=0.02)
mpimg.imsave(