def task_31():
gt_annot_file = 'datasets/ai_challenge_s03_c010-full_annotation.xml'
frames_path = 'datasets/AICity_data/train/S03/c010/data'
det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/our_results_finetune_faster.txt")
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
video_n_frames = number_of_images_jpg(frames_path)
det_bb_max_iou, idd = tracking_iou(frames_path, copy.deepcopy(det_bb), video_n_frames, mode='of')
ap_max_iou = calculate_ap(det_bb_max_iou, gt_bb, 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
det_bb_max_iou, idd = tracking_iou(frames_path, copy.deepcopy(det_bb), video_n_frames, mode='other')
ap_max_iou = calculate_ap(det_bb_max_iou, copy.deepcopy(gt_bb), 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
new_tracks = restore_tracks(frames_path, det_bb_max_iou)
ap_max_iou = calculate_ap(new_tracks, copy.deepcopy(gt_bb), 0, video_n_frames, mode='sort')
print("Ap after tracking with maximum IoU and of: {}".format(ap_max_iou))
ini_frame = 700
end_frame = 900
animation_tracks(new_tracks, idd, ini_frame, end_frame, frames_path)
def task1(gt_file):
gt = read_xml_gt(gt_file)
classes_to_keep = ['car']
gt = filter_gt(gt, classes_to_keep)
det_bb = generate_noisy_annotations(gt)
lst_gt = [item[0] for item in gt]
last_frame = np.max(lst_gt)
miou = 0
for f_val in range(0, last_frame):
frame_gt_bb = [gt[i] for i, num in enumerate(gt) if num[0] == f_val]
frame_det_bb = [det_bb[i] for i, num in enumerate(det_bb) if num[0] == f_val]
miou += frame_miou(frame_det_bb, frame_gt_bb, confidence=False)
miou = miou/last_frame
print("Noisy GT (Sorted Random): ", calculate_ap(det_bb, gt, 'random'))
print("Noisy GT (Sorted by Area): ", calculate_ap(det_bb, gt, 'area'))
preds_mask = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_mask_rcnn.txt")
print("maskrcnn ap: {}".format(calculate_ap(preds_mask, gt, 'sort')))
preds_ssd = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_ssd512.txt")
print("ssd ap: {}".format(calculate_ap(preds_ssd, gt, 'sort')))
preds_yolo = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_yolo3.txt")
print("yolo ap: {}".format(calculate_ap(preds_yolo, gt, 'sort')))
def task22(gt_annot_file, detections_file, frames_path):
ap_mode = 'area'
#Read and filter detections
det_bb = read_detections_file(detections_file)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
#Read and filter gt
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
#Calculate original ap
video_n_frames = number_of_images_jpg(frames_path)
# Constant velocity
det_bb_vel = kalman_filter_tracking(det_bb, video_n_frames, 0)
# Constant acceleration
det_bb_acc = kalman_filter_tracking(det_bb, video_n_frames, 1)
#Print ap results
print("Original ap: {}".format(
calculate_ap(det_bb, gt_bb, 0, video_n_frames, mode=ap_mode)))
print("Ap after tracking with constant velocity: {}".format(
calculate_ap(det_bb_vel, gt_bb, 0, video_n_frames, mode=ap_mode)))
print("Ap after tracking with constant acceleration: {}".format(
calculate_ap(det_bb_acc, gt_bb, 0, video_n_frames, mode=ap_mode)))
#Obtain animation
max_id_vel = max(det_bb_vel, key=lambda x: x[2])[2]
max_id_acc = max(det_bb_acc, key=lambda x: x[2])[2]
ini_frame = 550
end_frame = 650
def task21(gt_annot_file, detections_file, frames_path):
ap_mode = 'area'
# Read and filter detections
# det_bb = read_detections_file(detections_file)
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_mask_rcnn.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_ssd512.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/det_yolo3.txt")
# det_bb = read_detections_file("datasets/AICity_data/train/S03/c010/det/our_results_coco_faster.txt")
det_bb = read_detections_file(
"datasets/AICity_data/train/S03/c010/det/our_results_finetune_faster.txt"
)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
#Read and filter gt
gt_bb = read_xml_gt_options(gt_annot_file, False, False)
gt_bb = filter_gt(gt_bb, ["car"])
video_n_frames = number_of_images_jpg(frames_path)
original_ap = calculate_ap(det_bb, gt_bb, 0, video_n_frames, mode=ap_mode)
det_bb_max_iou, idd = tracking_iou(copy.deepcopy(det_bb), video_n_frames)
ap_max_iou = calculate_ap(det_bb_max_iou,
gt_bb,
0,
video_n_frames,
mode=ap_mode)
print("Original ap: {}".format(original_ap))
print("Ap after tracking with maximum IoU: {}".format(ap_max_iou))
ini_frame = 550
end_frame = 650
animation_tracks(det_bb_max_iou, idd, ini_frame, end_frame, frames_path)
def task11(images_path, gt_annot_file, config_file):
files = get_files_from_dir(images_path, 'jpg')
# we remove bikes and compute ap only for car class
gt_bb = read_xml_gt(gt_annot_file)
classes_to_keep = ['car']
gt_bb = filter_gt(gt_bb, classes_to_keep)
filename = "our_results_coco_{}.txt".format("faster" if "faster" in
config_file else "retina")
preds = inference(config_file,
files,
save_results=True,
out_filename=filename)
ap50 = calculate_ap(preds, gt_bb, 0, len(files), mode='area')
print(ap50)
filename_gif = "faster_on_coco" if "faster" in config_file else "retina_on_coco"
animation_2bb(
filename_gif,
'.gif',
gt_bb,
preds,
images_path,
ini=800,
)
def task1(frames_path,
detections_file,
gt_file,
tracking_method,
postprocessing=True,
visualization=False):
print("Start loading data")
# Read gt
gt_bb = read_gt_txt(gt_file)
# Read and filter detections
det_bb = read_detections_file(detections_file)
det_bb = filter_det_confidence(det_bb, threshold=0.5)
# Tracking
print("Start tracking")
video_n_frames = number_of_images_jpg(frames_path)
if tracking_method == "MaxOverlap":
det_bb_tracking, id_max = tracking_iou(frames_path,
copy.deepcopy(det_bb),
video_n_frames,
mode='other')
elif tracking_method == "Kalman":
det_bb_tracking = kalman_filter_tracking(copy.deepcopy(det_bb),
video_n_frames, 0)
id_max = max([v[2] for v in det_bb_tracking])
else:
raise NotImplemented
# Postprocessing
if postprocessing:
print("Starting postprocessing tracking")
det_bb_tracking = clean_tracks(copy.deepcopy(det_bb_tracking), id_max)
det_bb_tracking = remove_parked(copy.deepcopy(det_bb_tracking),
id_max,
threshold=10.0)
# Results
print("Start obtaining metrics")
ap = calculate_ap(det_bb_tracking,
copy.deepcopy(gt_bb),
0,
video_n_frames,
mode='sort')
print("Average precision: {}".format(ap))
# Create animation if required
if visualization:
print("Start storing animation")
ini_frame = 700
end_frame = 800
frames_path = 'datasets/AICity_data/train/S03/c010/data'
animation_tracks(det_bb_tracking, id_max, ini_frame, end_frame,
frames_path)
return det_bb_tracking
def task12(images_path, config_file, dataset_annot_file, gt_annot_file):
gt_bb = read_xml_gt(gt_annot_file)
classes_to_keep = ['car']
gt_bb = filter_gt(gt_bb, classes_to_keep)
filename = "our_results_finetune_{}.txt".format("faster" if "faster" in
config_file else "retina")
det_bb = train(config_file,
images_path,
dataset_annot_file,
out_filename=filename)
ap50 = calculate_ap(det_bb, gt_bb, 0, 2140, mode='area')
print(ap50)
filename_gif = "faster_finetune" if "faster" in config_file else "retina_finetune"
animation_2bb(
filename_gif,
'.gif',
gt_bb,
det_bb,
images_path,
ini=800,
)
path = 'C:/Users/Sara/Documents/GitHub/Visual-recognition/week6/testing_sara/KITTI-MOTS/instances_txt/0014.txt'
gt_bb = read_gt_txt(path, mode='car')
det_bb = read_det(cat='car')
frame_path = "C:/Users/Sara/Documents/GitHub/Visual-recognition/week6/testing_sara/KITTI-MOTS/training/image_02/0014"
video_n_frames = number_of_images(frame_path)
det_bb_max_iou, idd = tracking_iou(frame_path,
copy.deepcopy(det_bb),
video_n_frames,
mode='normal')
ap_max_iou = calculate_ap(det_bb_max_iou,
copy.deepcopy(gt_bb),
0,
video_n_frames,
mode='area')
frame_with_2bb(gt_bb, det_bb, frame_path, 10)
animation_tracks(det_bb_max_iou, idd, 0, video_n_frames, frame_path)
f = open('0014.txt', 'w')
for ele in det_bb_max_iou:
f.write(str(ele[0]) + ' ')
f.write(str(ele[2]) + ' ')
f.write(str(ele[1]) + ' ')
f.write(str(ele[8]) + ' ')
f.write(str(ele[9]) + ' ')
f.write(str(ele[10]) + '\n')
def task1(gt_path, color_space=cv2.COLOR_BGR2GRAY):
path = 'datasets/AICity_data/train/S03/c010'
save_frames(path)
print("Finished saving")
frames_path = 'datasets/AICity_data/train/S03/c010/data'
video_n_frames = number_of_images_jpg(frames_path)
# this is very time consuming, we should avoid comuting it more than once.
mu, sigma = bg_model(frames_path, color_space)
# Gaussian plot in the slides
xx = np.linspace(0, 255, 1000)
exp = 1 / (sigma[561, 712] * np.sqrt(2 * np.pi)) * \
np.exp(-(xx - mu[561, 712]) ** 2 / (2 * sigma[561, 712] ** 2))
plt.figure()
plt.plot(xx, exp, color='blue', label='Distribution')
plt.axvline(x=mu[561, 712], color='red', label='Mean')
plt.axvline(x=mu[561, 712] - (sigma[561, 712] + 2),
linestyle='--',
color='green',
label='Tolerance ' + r'$\alpha = 1$')
plt.axvline(x=mu[561, 712] + (sigma[561, 712] + 2),
linestyle='--',
color='green')
plt.axvline(x=mu[561, 712] - 3 * (sigma[561, 712] + 2),
linestyle='--',
color='purple',
label='Tolerance ' + r'$\alpha = 5$')
plt.axvline(x=mu[561, 712] + 3 * (sigma[561, 712] + 2),
linestyle='--',
color='purple')
plt.legend()
alphas = [1, 1.5, 2, 2.5, 3, 3.5, 4, 4.5, 5, 5.5, 6, 6.5, 7]
aps7 = []
# Test for different alpha values
for a in alphas:
det_bb = remove_bg(mu,
sigma,
a,
frames_path,
int(video_n_frames * 0.25),
video_n_frames,
animation=False,
denoise=True,
adaptive=False,
color_space=color_space)
gt_bb = read_xml_gt_options(gt_path, True, True)
ap = calculate_ap(det_bb,
gt_bb,
int(video_n_frames * 0.25),
video_n_frames,
mode='area')
animation_2bb('try_dnoise', '.gif', gt_bb, det_bb, frames_path, 10, 10,
int(video_n_frames * 0.25), int(1920 / 4), int(1080 / 4))
print(a, ap)
aps7.append(ap)
plt.title('Median Filter')
# plt.plot(alphas, aps3, label = 'Window size 3')
# plt.plot(alphas, aps5, label = 'Window size 5')
plt.plot(alphas, aps7, label='Window size 7')
plt.xlabel(r'$\alpha$')
plt.ylabel('mAP')
plt.legend()
def task3(frames_path,
annots_file,
estimation_percent=.25,
save_to_disk=False,
save_size=(480, 270)):
ims = get_files_from_dir(frames_path, "jpg")
gt_bb = read_xml_gt_options(annots_file, True, True)
save_to_disk = False
init_images = int(len(ims) * estimation_percent)
first_eval_image = init_images + 1
modes = ['mog', 'knn', 'gmg', 'LSBP']
model_params = {
'mog': {
# 'varThreshold': 32
},
'knn': {
# 'dist2Threshold': 350
},
'gmg': {
# 'initializationFrames': init_images
},
'LSBP': {},
}
for mode in modes:
bg_sub = bg_estimation(mode, **model_params[mode])
det_bb = []
if save_to_disk:
frames = np.zeros(
(len(ims) - first_eval_image, save_size[1], save_size[0]))
for i, im in enumerate(ims):
im = cv2.imread(im, 0)
fg = bg_sub.apply(im)
fg[fg != 255] = 0
# fg = denoise_bg(fg)
if i >= first_eval_image:
det_bb += fg_segmentation_to_boxes(fg, i)
if save_to_disk:
fg = cv2.resize(fg, save_size)
frames[i - first_eval_image, ...] = fg
ap = calculate_ap(det_bb,
gt_bb,
init_images + 1,
len(ims),
mode='area')
if save_to_disk:
animation_2bb(
'{}_bg'.format(mode),
'.gif',
gt_bb,
det_bb,
frames_path,
ini=800,
)
frames_to_gif("fg_{}.gif".format(mode), frames[:100])
print("{} - {}".format(mode, ap))
def task2(frames_path, gt_path, color_space=cv2.COLOR_BGR2GRAY, channels=(0)):
grid_search = False
save_videos = False
fine_tune_search = False
videos_rgb_bb = True
gt_bb = read_xml_gt_options(gt_path, True, True)
classes_to_keep = ['car', 'bike']
gt_bb = filter_gt(gt_bb, classes_to_keep)
video_n_frames = number_of_images_jpg(frames_path)
mu, sigma = bg_model(frames_path, color_space)
if grid_search:
mAPs = []
alphas = [2, 2.5, 3, 3.5, 4, 4.5]
rhos = [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
for alpha in alphas:
mAP_same_alfa = []
for rho in rhos:
det_bb = remove_bg(mu,
sigma,
alpha,
frames_path,
int(video_n_frames * 0.25),
video_n_frames,
color_space=color_space,
adaptive=True,
rho=rho,
channels=channels)
mAP = calculate_ap(det_bb,
gt_bb,
int(video_n_frames * 0.25),
video_n_frames,
mode='area')
mAP_same_alfa.append(mAP)
print("Alpha: {:2f} | Rho: {:2f} | mAP: {:2f} |".format(
alpha, rho, mAP))
mAPs.append(mAP_same_alfa)
# Plot the surface
X, Y = np.meshgrid(alphas, rhos)
fig = plt.figure()
ax = fig.gca(projection='3d')
surf = ax.plot_surface(X,
Y,
mAPs,
rstride=0.1,
cstride=0.1,
cmap='viridis',
edgecolor='none')
ax.set_xlabel('Alpha')
ax.set_ylabel('Rho')
ax.set_zlabel('mAP')
# Add a color bar which maps values to colors.
fig.colorbar(surf, shrink=0.5, aspect=5)
plt.savefig("_grid_search.png")
plt.show()
if save_videos:
alpha = 3.2
rho = 0.02
det_bb = remove_bg(mu,
sigma,
alpha,
frames_path,
800,
900,
rho=rho,
denoise=True,
animation=True,
color_space=color_space,
adaptive=True,
channels=channels)
if fine_tune_search:
mAPs = []
alphas = [
3, 3, 3, 3.5, 3.5, 3.5, 4, 4, 4, 3.5, 3.5, 3.5, 3.5, 3.7, 3.2, 3.1,
3.3
]
rhos = [
0.0, 0.2, 0.4, 0.0, 0.2, 0.4, 0.0, 0.2, 0.4, 0.1, 0.3, 0.15, 0.05,
0.03, 0.02, 0.01, 0.01
]
for i in range(0, len(alphas)):
det_bb = remove_bg(mu,
sigma,
alphas[i],
frames_path,
int(video_n_frames * 0.25),
video_n_frames,
color_space=color_space,
adaptive=True,
rho=rhos[i],
denoise=True)
mAP = calculate_ap(det_bb,
gt_bb,
int(video_n_frames * 0.25),
video_n_frames,
mode='area')
print("Alpha: {:2f} | Rho: {:2f} | mAP: {:2f} |".format(
alphas[i], rhos[i], mAP))
mAPs.append(mAP)
fig, ax = plt.subplots()
plt.xlim(2.9, 4.1)
plt.ylim(-0.1, 0.5)
ax.set_aspect(1)
ax.set_xlabel('Alpha')
ax.set_ylabel('Rho')
for i in range(0, len(mAPs)):
circle = plt.Circle((alphas[i], rhos[i]), mAPs[i] / 25)
ax.add_artist(circle)
plt.savefig("fine_tune_search.png", bbox_inches='tight')
plt.show()
if videos_rgb_bb:
alpha = 2.5
rho = 0.02
det_bb = remove_bg(mu,
sigma,
alpha,
frames_path,
int(video_n_frames * 0.25),
video_n_frames,
color_space=color_space,
adaptive=True,
rho=rho,
denoise=True,
channels=channels)
print("Calculating AP")
mAP = calculate_ap(det_bb,
gt_bb,
int(video_n_frames * 0.25),
video_n_frames,
mode='area')
animation_2bb('rgb_bb_adaptive', '.gif', gt_bb, det_bb, frames_path,
10, 10, 800, int(1920 / 4), int(1080 / 4))
print(f"Channels: {channels}, Colorspace: {color_space}, mAP: {mAP}")
