# Iterate over testing frames to choose one with bounding boxes
for test_frame in test_frames:
# Get frame ID from frame filename
frm = ut.frameIdfrom_filename(test_frame)
# Get mask and list of bounding boxes from the
_, cbbox = ut.getbboxmask(bboxes_gt, frm, (s[0], s[1]))
# get current frame in the experiment format
I = ut.getImg_D(test_frame,
D=DIM,
color_space=COLOR_SPACE,
color_channels=COLOR_CHANNELS)
# get the mask
map_bg = bg.foreground_from_GBGmodel(mu_bg,
std_bg,
test_frame,
th=alpha)
# refine mask with morphological filter
if F_MORPH:
kernel = np.ones((11, 11))
morp_masks = binary_fill_holes(closing(map_bg, kernel))
morp_masks = binary_fill_holes(opening(morp_masks, kernel))
map_bg = morp_masks
# Get BBox from mask
bboxes_in_img = bg.connected_components(map_bg,
area_min=AREA_MIN,
area_max=AREA_MAX,
ff_min=FF_MIN,
def main():
"""
Function to compute the background and foreground of a sequence
of frames using a Gaussian distribution method.
This script uses the 100*N % first frames for training and the
rest for testing.
Meaningful variables are defined at the beginning of this script.
"""
# Set useful directories
frames_dir = os.path.join(src.ROOT_DIR, 'datasets', 'm6_week1_frames',
'frames')
results_dir = os.path.join(src.OUTPUT_DIR, WEEK, TASK, EXP_NAME)
# Ground truth file path
gt_file = os.path.join(src.ROOT_DIR, 'datasets', 'AICity_data', 'train',
'S03', 'c010', 'gt', 'gt.txt')
# Create folders if they don't exist
if not os.path.isdir(results_dir):
os.mkdir(results_dir)
# Get file paths for each of the frames and sort them according
# to the frame number
frame_paths = ut.get_files_from_dir2(frames_dir, ext='.jpg')
frame_paths.sort(key=ut.natural_keys)
# Total number of frames
num_frames = len(frame_paths)
# Flag to show the results based on image dimension
color_flag = cv.IMREAD_GRAYSCALE if DIM <= 1 else cv.IMREAD_COLOR
# Get the the images for training
num_frames_test = int(num_frames * N)
# Separate frames for training and testing
train_frames = frame_paths[:num_frames_test]
test_frames = frame_paths[num_frames_test:]
# Print useful information
print("Total number of frames : {}".format(num_frames))
print("Number of frames for training: {}".format(len(train_frames)))
print("Number of frames for testing : {}".format(len(test_frames)))
# Model numpy files
mu_file = os.path.join(results_dir, 'mu.npy')
std_file = os.path.join(results_dir, 'std.npy')
# If the files exist, load the values. If not, compute them
if os.path.isfile(mu_file):
mu_bg = np.load(mu_file)
std_bg = np.load(std_file)
else:
mu_bg, std_bg = bg.getGauss_bg(train_frames, D=DIM, gt_file=gt_file)
# Plot the mean and the standard deviation computed and save it
fig = plt.figure(1, figsize=(6, 8))
ax1 = plt.subplot(211)
ax2 = plt.subplot(212)
if DIM == 3:
ax1.imshow(mu_bg, vmin=0, vmax=255)
ax2.imshow(std_bg, vmin=0, vmax=255)
else:
ax1.imshow(mu_bg, cmap='gray')
ax2.imshow(std_bg, cmap='gray')
ax1.set_title("Mean background model over {} frames".format(
len(train_frames)))
ax2.set_title("Standard noise backgrond model")
plt.savefig(os.path.join(results_dir, 'mean_std_testing.png'))
# Threshold to create different masks
alphas = np.linspace(1.5, 0.25, 4.0)
# Get image size of the frames
frame_img = cv.imread(test_frames[1], color_flag)
# Get bounding boxes from ground truth
bboxes_gt = ut.get_bboxes_from_MOTChallenge(gt_file)
# Iterate over testing frames to choose one with bounding boxes
for test_frame in test_frames:
# Get frame ID from frame filename
frm = ut.frameIdfrom_filename(test_frame)
# TODO: VER
# Get mask and list of bounding boxes from the
fore_mask, cbbox = ut.getbboxmask(bboxes_gt, frm, frame_img.shape[:2])
# If there are bounding boxes in the ground truth
if any(cbbox):
frame_img = cv.imread(test_frame, color_flag)
break
# Plot different thresholds
fig, axs = plt.subplots(2,
3,
figsize=(15, 6),
facecolor='w',
edgecolor='g')
fig.subplots_adjust(hspace=.5, wspace=.01)
axs = axs.ravel()
axs[0].imshow(frame_img, cmap='gray')
axs[0].set_title(os.path.basename(test_frame))
# Iterate over each bounding box in the ground truth and add them
# to the image
for bbox in cbbox:
# Draw rectangle in the image
rect = patches.Rectangle((bbox[0], bbox[2]),
bbox[1] - bbox[0],
bbox[3] - bbox[2],
linewidth=1,
edgecolor='r',
facecolor='none')
# Add the patch to the Axes
axs[0].add_patch(rect)
axs[1].imshow(fore_mask, cmap='gray')
axs[1].set_title('GT map')
# Plot the mask for the different thresholds
i = 2
for alpha in alphas:
map_bg = bg.foreground_from_GBGmodel(mu_bg,
std_bg,
frame_img,
th=alpha)
axs[i].imshow(map_bg, cmap='gray')
axs[i].set_title("th = {}".format(alpha))
i += 1
plt.savefig(os.path.join(results_dir, 'thresholds.png'))
print("Done!")
if d==2:
s = np.shape(I)
I2 = np.dstack((I,np.zeros((s[0],s[1]))))
axs[0].imshow(I2,vmin=0,vmax=255)
#ax2.imshow(stdBG2,vmin=0,vmax=255)
else:
axs[0].imshow(I,cmap='gray')
axs[0].set_title(testing_list[loc])
for b in cbbox:
rect = patches.Rectangle((b[0],b[2]),b[1]-b[0],b[3]-b[2],linewidth=1,edgecolor='r',facecolor='none')
# Add the patch to the Axes
axs[0].add_patch(rect)
axs[1].imshow(m0,cmap='gray')
axs[1].set_title('GT map')
i= 2
for a in th:
mapBG = bg.foreground_from_GBGmodel(muBG,stdBG,I,th =a)
axs[i].imshow(mapBG,cmap='gray')
axs[i].set_title("th={}".format(a))
i+=1
plt.show()
# showing the background model , mu +noise
#print("mu: {},var: {}".format(len(muBG),len(varBG)))
color_space=COLOR_SPACE,
color_channels=COLOR_CHANNELS)
if d == 1:
I = np.squeeze(I, axis=2)
I = np.squeeze(I, axis=2)
frm = ut.frameIdfrom_filename(
testing_list[loc]) # Get the number of the frame from filename
print(frm)
_, cbbox = ut.getbboxmask(
Bbox, frm, (s[0], s[1])) # List of GT Bboxes per specific frame
#m0, cbbox = ut.getbboxmask(Bbox, frm, (s[0], s[1]))
mapBG = bg.foreground_from_GBGmodel(
muBG, stdBG, I, th=11) # Mask of foreground for specific frame
#plt.imshow(mapBG)
#plt.show()
# Refine masks and obtain Bboxes:
# If morph mask is set:
if F_MORPH:
#kernel = np.ones((3, 3))
kernel = np.ones((11, 11))
morp_masks = binary_fill_holes(closing(mapBG, kernel))
morp_masks = binary_fill_holes(opening(morp_masks, kernel))
# Iterate over testing frames to choose one with bounding boxes
for test_frame in test_frames:
# Get frame ID from frame filename
frm = ut.frameIdfrom_filename(test_frame)
# Get mask and list of bounding boxes from the
_, cbbox = ut.getbboxmask(bboxes_gt, frm, (s[0],s[1]))
# get current frame in the experiment format
I = ut.getImg_D(test_frame,D=DIM,color_space = COLOR_SPACE,color_channels= COLOR_CHANNELS)
if DIM ==1:
I=np.squeeze(I,axis=2)
# get the mask
map_bg,mu_bg, std_bg = bg.get_BG(mu_bg, std_bg,I,th=alpha,adaptive =ADAPTIVE, p=p)
map_bg = bg.foreground_from_GBGmodel(mu_bg, std_bg, I, th=alpha)
# refine mask with morphological filter
if F_MORPH:
kernel = np.ones((11, 11))
morp_masks = binary_fill_holes(closing(map_bg, kernel))
morp_masks = binary_fill_holes(opening(morp_masks, kernel))
map_bg = morp_masks
# Get BBox from mask
bboxes_in_img = bg.connected_components(map_bg, area_min=AREA_MIN, area_max=AREA_MAX,
ff_min=FF_MIN, ff_max=FF_MAX, fr_min=FR_MIN, plot=PLOT_BBOX)
# Get mesurments for current frame
"""
GOOD IDEA IS TO SAVE TO BG MODEL SO WE CAN RUN ON IT DIFFERENT TESTS
"""
fig = plt.figure(1)
ax1 = plt.subplot(211)
#ax3 = plt.subplot(312)
ax2 = plt.subplot(212)
ax1.imshow(muBG2, cmap='gray')
ax2.imshow(stdBG2, cmap='gray')
ax1.set_title("Mean BG model - #" + str(len(trainig_list)) + " frames")
ax2.set_title("Std noise BG model")
plt.show()
# Testing Example
th = 3
I = cv.imread(testing_list[500], Clr_flag)
mapBG = bg.foreground_from_GBGmodel(muBG2, stdBG2, I, th=th)
fig = plt.figure(1)
ax1 = plt.subplot(211)
ax2 = plt.subplot(212)
ax1.imshow(I, cmap='gray')
ax2.imshow(mapBG, cmap='gray')
ax1.set_title(testing_list[500])
ax2.set_title("Foreground Map, with th=" + str(th))
plt.show()
# showing the background model , mu +noise
#print("mu: {},var: {}".format(len(muBG),len(varBG)))