def create_dataset(videos_list=None,
savepath=joints_path(),
im_regularizer=reg.Regularizer(),
heat_regularizer=reg.Regularizer(),
fillgaps=False,
cross_radius=3,
enlarge=0.2,
shade=False):
"""reads the videos specified as parameter and for each frame produces and saves a .mat file containing
the frame, the corresponding heatmap indicating the position of the hand and the modified depth.
:param fillgaps: set to True to also get interpolated frames
:param im_regularizer: object used to regularize the images
:param heat_regularizer: object used to regularize the heatmaps
:param savepath: path of the folder where the produces .mat files will be saved. If left to the default value None,
the /resources/hands_bounding_dataset/hands_rgbd_transformed folder will be used
:param videos_list: list of videos you need the .mat files of. If left to the default value None, all videos will
be exploited
:param cross_radius: radius of the crosses of the heatmaps
:param enlarge: crops enlarge factor
:param shade: set to true to shade the pixels that identify a junction in a heatmap according to their
distance with the center (real position of junction"""
if savepath is None:
basedir = joints_path()
else:
basedir = savepath
if not os.path.exists(basedir):
os.makedirs(basedir)
framesdir = resources_path("framedata")
if videos_list is None:
vids = os.listdir(framesdir)
vids = [x for x in vids if os.path.isdir(os.path.join(framesdir, x))]
else:
vids = videos_list
for vid in tqdm.tqdm(vids):
frames, labels = load_labelled_videos(vid, fillgaps=fillgaps)
# depths, _ = load_labelled_videos(vid, getdepth=True, fillgaps=fillgaps)
fr_num = frames.shape[0]
for i in tqdm.tqdm(range(0, fr_num)):
if labels[i] is not None:
try:
frame = frames[i]
label = labels[i][:, 0:2]
visible = labels[i][:, 2:3]
label *= [frame.shape[1], frame.shape[0]]
label = np.array(label, dtype=np.int32).tolist()
label = [[p[1], p[0]] for p in label]
coords = __get_coord_from_labels(label)
cut = u.crop_from_coords(frame, coords, enlarge)
heatmaps = __create_21_heatmaps(label, coords,
np.shape(frame),
cross_radius, enlarge,
shade)
cut = im_regularizer.apply(cut)
heatmaps = heat_regularizer.apply_on_batch(heatmaps)
# heatmaps = __heatmaps_dim_reducer(heatmaps)
heatmaps = __stack_heatmaps(heatmaps)
path = os.path.join(basedir, vid + "_" + str(i))
__persist_frame(path, cut, heatmaps, visible)
except ValueError as e:
print("Error " + e + " on vid " + vid + str(i))
def create_dataset(videos_list=None,
savepath=None,
im_regularizer=reg.Regularizer(),
fillgaps=False,
enlarge=0.5):
"""reads the videos specified as parameter and for each frame produces and saves a .mat file containing
the frame, the corresponding heatmap indicating the position of the hand and the modified depth.
:param fillgaps: set to True to also get interpolated frames
:param im_regularizer: object used to regularize the cuts
:param savepath: path of the folder where the produces .mat files will be saved. If left to the default value None,
the /resources/hands_bounding_dataset/hands_rgbd_transformed folder will be used
:param videos_list: list of videos you need the .mat files of. If left to the default value None, all videos will
be exploited
:param enlarge: crops enlarge factor"""
if savepath is None:
basedir = resources_path("palm_back_classification_dataset")
else:
basedir = savepath
if not os.path.exists(basedir):
os.makedirs(basedir)
framesdir = resources_path("framedata")
if videos_list is None:
vids = os.listdir(framesdir)
vids = [x for x in vids if os.path.isdir(os.path.join(framesdir, x))]
else:
vids = videos_list
for vid in tqdm.tqdm(vids):
frames, labels = load_labelled_videos(vid, fillgaps=fillgaps)
fr_num = frames.shape[0]
lr = get_right_left(vid)
if lr == LEFT:
lr = -1
for i in tqdm.tqdm(range(0, fr_num)):
if labels[i] is not None:
try:
frame = frames[i]
label = labels[i][:, 0:2]
# conf is a real in [-1.0, 1.0] such that -1.0 is full back, +1.0 is full palm
# middle values express partial confidence, but all info is in one single value
conf = pb.leftright_to_palmback(
hand=hand_format(label),
side=pb.RIGHT if lr == RIGHT else pb.LEFT)
# if you want the crisp result, there it is:
result = PALM if conf >= 0 else BACK
if result < 0:
result = 0
# and the confidence on that result is in [0..1]:
conf = abs(conf)
label *= [frame.shape[1], frame.shape[0]]
label = np.array(label, dtype=np.int32).tolist()
label = [[p[1], p[0]] for p in label]
coords = __get_coord_from_labels(label)
cut = u.crop_from_coords(frame, coords, enlarge)
cut = im_regularizer.apply(cut)
path = os.path.join(basedir, vid + "_" + str(i))
__persist_frame(path, cut, result, conf)
except ValueError as e:
print("Error " + str(e) + " on vid " + vid + str(i))
def create_dataset(videos_list=None,
savepath=None,
im_regularizer=reg.Regularizer(),
fillgaps=False,
enlarge=0.5,
data_augment=False):
"""reads the videos specified as parameter and for each frame produces and saves a .mat file containing
the frame, the corresponding heatmap indicating the position of the hand and the modified depth.
:param fillgaps: set to True to also get interpolated frames
:param im_regularizer: object used to regularize the cuts
:param savepath: path of the folder where the produces .mat files will be saved. If left to the default value None,
the /resources/hands_bounding_dataset/hands_rgbd_transformed folder will be used
:param videos_list: list of videos you need the .mat files of. If left to the default value None, all videos will
be exploited
:param enlarge: crops enlarge factor"""
if savepath is None:
basedir = resources_path("left_right_classification_dataset")
else:
basedir = savepath
if not os.path.exists(basedir):
os.makedirs(basedir)
framesdir = resources_path("framedata")
if videos_list is None:
vids = os.listdir(framesdir)
vids = [x for x in vids if os.path.isdir(os.path.join(framesdir, x))]
else:
vids = videos_list
for vid in tqdm.tqdm(vids):
frames, labels = load_labelled_videos(vid, fillgaps=fillgaps)
fr_num = frames.shape[0]
result = get_right_left(vid)
for i in tqdm.tqdm(range(0, fr_num)):
if labels[i] is not None:
try:
frame = frames[i]
label = labels[i][:, 0:2]
label *= [frame.shape[1], frame.shape[0]]
label = np.array(label, dtype=np.int32).tolist()
label = [[p[1], p[0]] for p in label]
coords = __get_coord_from_labels(label)
cut = u.crop_from_coords(frame, coords, enlarge)
cut = im_regularizer.apply(cut)
path = os.path.join(basedir, vid + "_" + str(i))
__persist_frame(path, cut, result)
if data_augment:
path = os.path.join(basedir, vid + "_t_" + str(i))
transp = cut.squeeze().transpose()
transp = transp.reshape(
[transp.shape[0], transp.shape[1], 1])
__persist_frame(path, transp, 1 - result)
except ValueError as e:
print("Error " + str(e) + " on vid " + vid + str(i))
def __create_21_heatmaps(label,
coords_for_cut,
original_shape,
cross_radius,
enlarge,
shade=False):
heatmaps = []
for l in label:
heat = np.zeros([original_shape[0], original_shape[1]])
heat = __set_cross(heat, l, cross_radius, shade)
heat = u.crop_from_coords(heat, coords_for_cut, enlarge)
heatmaps.append(heat)
return np.array(heatmaps)