# label_text: group label text in UI
# choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# value: symbolic value name for choice
# label_text: choice label text in UI
# tooltip: extra information for user
ANNO_CONTROLS = [
labelling_tool.AnnoControlCheckbox('good_quality', 'Good quality'),
labelling_tool.AnnoControlRadioButtons(
'visibility',
'Visible',
choices=[
labelling_tool.AnnoControlRadioButtons.choice(
value='full',
label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='mostly',
label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='obscured',
label_text='Obscured',
tooltip='Object is significantly obscured'),
],
label_on_own_line=False),
labelling_tool.AnnoControlPopupMenu(
'material',
'Material',
groups=[
labelling_tool.AnnoControlPopupMenu.group(
label_text='Artifical/buildings',
choices=[
def run_app(images_dir, images_pat, labels_dir, readonly,
update_label_object_ids, enable_dextr, dextr_weights):
if enable_dextr or dextr_weights is not None:
from dextr.model import DextrModel
import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if dextr_weights is not None:
dextr_weights = pathlib.Path(dextr_weights).expanduser()
dextr_model = torch.load(dextr_weights, map_location=device)
else:
dextr_model = DextrModel.pascalvoc_resunet101().to(device)
dextr_model.eval()
dextr_fn = lambda image, points: dextr_model.predict([image], points[
None, :, :])[0] >= 0.5
else:
dextr_fn = None
# Load schema
schema_path = pathlib.Path(images_dir) / 'schema.json'
schema_store = labelling_schema.FileSchemaStore(schema_path,
readonly=readonly)
# Annotation controls
# Labels may also have optional meta-data associated with them
# You could use this for e.g. indicating if an object is fully visible, mostly visible or significantly obscured.
# You could also indicate quality (e.g. blurriness, etc)
# There are four types of annotation. They have some common properties:
# - name: symbolic name (Python identifier)
# - label_text: label text in UI
# Check boxes, radio buttons and popup menus also have:
# - visibility_label_text: [optional] if provided, label visibility can be filtered by this annotation value,
# in which case a drop down will appear in the UI allowing the user to select a filter value
# that will hide/show labels accordingly
# Control types:
# Check box (boolean value):
# `labelling_tool.AnnoControlCheckbox`; only the 3 common parameters listed above
# Radio button (choice from a list):
# `labelling_tool.AnnoControlRadioButtons`; the 3 common parameters listed above and:
# - choices: list of `labelling_tool.AnnoControlRadioButtons.choice` that provide:
# - value: symbolic value name for choice
# - tooltip: extra information for user
# - label_on_own_line [optional]: if True, place the label and the buttons on a separate line in the UI
# Popup menu (choice from a grouped list):
# `labelling_tool.AnnoControlPopupMenu`; the 3 common parameters listed above and:
# - groups: list of groups `labelling_tool.AnnoControlPopupMenu.group`:
# - label_text: label text in UI
# - choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# - value: symbolic value name for choice
# - label_text: choice label text in UI
# - tooltip: extra information for user
# Text (free form plain text):
# `labelling_tool.AnnoControlText`; only the 2 common parameters listed above and:
# - multiline: boolean; if True a text area will be used, if False a single line text entry
anno_controls = [
labelling_tool.AnnoControlCheckbox(
'good_quality',
'Good quality',
visibility_label_text='Filter by good quality'),
labelling_tool.AnnoControlRadioButtons(
'visibility',
'Visible',
choices=[
labelling_tool.AnnoControlRadioButtons.choice(
value='full',
label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='mostly',
label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='obscured',
label_text='Obscured',
tooltip='Object is significantly obscured'),
],
label_on_own_line=False,
visibility_label_text='Filter by visibility'),
labelling_tool.AnnoControlPopupMenu(
'material',
'Material',
groups=[
labelling_tool.AnnoControlPopupMenu.group(
label_text='Artifical/buildings',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='concrete',
label_text='Concrete',
tooltip='Concrete objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='plastic',
label_text='Plastic',
tooltip='Plastic objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='asphalt',
label_text='Asphalt',
tooltip='Road, pavement, etc.'),
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Flat natural',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='grass',
label_text='Grass',
tooltip='Grass covered ground'),
labelling_tool.AnnoControlPopupMenu.choice(
value='water',
label_text='Water',
tooltip='Water/lake')
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Vegetation',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='trees', label_text='Trees',
tooltip='Trees'),
labelling_tool.AnnoControlPopupMenu.choice(
value='shrubbery',
label_text='Shrubs',
tooltip='Shrubs/bushes'),
labelling_tool.AnnoControlPopupMenu.choice(
value='flowers',
label_text='Flowers',
tooltip='Flowers'),
labelling_tool.AnnoControlPopupMenu.choice(
value='ivy', label_text='Ivy', tooltip='Ivy')
]),
],
visibility_label_text='Filter by material'),
# labelling_tool.AnnoControlText('comment', 'Comment', multiline=False),
]
image_pats = images_pat.split('|')
# Load in .JPG images from the 'images' directory.
labelled_images = labelled_image.LabelledImage.for_directory(
images_dir, image_filename_patterns=image_pats, readonly=readonly)
print('Loaded {0} images'.format(len(labelled_images)))
if update_label_object_ids:
n_updated = 0
for limg in labelled_images:
if limg.labels_store.labels_path.exists():
label_js = json.load(limg.labels_store.labels_path.open('r'))
prefix = str(uuid.uuid4())
modified = labelling_tool.ensure_json_object_ids_have_prefix(
label_js, id_prefix=prefix)
if modified:
with open(limg.labels_store.labels_path, 'w') as f_out:
json.dump(label_js, f_out, indent=3)
n_updated += 1
print('Updated object IDs in {} files'.format(n_updated))
# For documentation of the configuration, please see the comment above `labelling_tool.DEFAULT_CONFIG`
config = labelling_tool.DEFAULT_CONFIG
tasks = [
dict(name='finished', human_name='[old] finished'),
dict(name='segmentation', human_name='Outlines'),
dict(name='classification', human_name='Classification'),
]
flask_labeller_and_schema_editor(labelled_images,
schema_store,
tasks=tasks,
anno_controls=anno_controls,
config=config,
dextr_fn=dextr_fn)
def run_app(dextr_weights, enable_firebug, use_http_memory_cache):
import pathlib
import glob
from image_labelling_tool import labelled_image, labelling_tool
# Create an application
app = QtWidgets.QApplication([])
# Start with a dialog that allows the user to choose the images directory, optionally
# a labels directory and read only checkbox
init_dialog = QtWidgets.QDialog()
# Dialog layout
dia_layout = QtWidgets.QGridLayout()
init_dialog.setLayout(dia_layout)
images_dir = ['.']
labels_dir = [None]
# Prompt
prompt_label = QtWidgets.QLabel(
'Choose an images directory and optionally a labels directory.')
dia_layout.addWidget(prompt_label, 0, 0, 1, 2)
# Images directory
images_dir_button = QtWidgets.QPushButton('Choose images directory')
images_dir_button.setIcon(init_dialog.style().standardIcon(
QtWidgets.QStyle.SP_DirOpenIcon))
images_dir_label = QtWidgets.QLabel(images_dir[0])
def _on_images_dir_button():
file_dialog = QtWidgets.QFileDialog(None, 'Choose images directory',
'')
file_dialog.setFileMode(QtWidgets.QFileDialog.Directory)
choice = file_dialog.exec()
if choice == QtWidgets.QFileDialog.Accepted:
images_dir[0] = file_dialog.selectedFiles()[0]
images_dir_label.setText(images_dir[0])
images_dir_button.clicked.connect(_on_images_dir_button)
dia_layout.addWidget(images_dir_button, 1, 0)
dia_layout.addWidget(images_dir_label, 1, 1)
# Labels directory
labels_dir_button = QtWidgets.QPushButton(
'[Optional] Choose labels directory')
labels_dir_button.setIcon(init_dialog.style().standardIcon(
QtWidgets.QStyle.SP_DirOpenIcon))
labels_dir_label = QtWidgets.QLabel('')
def _on_labels_dir_button():
file_dialog = QtWidgets.QFileDialog(None, 'Choose labels directory',
'')
file_dialog.setFileMode(QtWidgets.QFileDialog.Directory)
choice = file_dialog.exec()
if choice == QtWidgets.QFileDialog.Accepted:
labels_dir[0] = file_dialog.selectedFiles()[0]
labels_dir_label.setText(labels_dir[0])
labels_dir_button.clicked.connect(_on_labels_dir_button)
dia_layout.addWidget(labels_dir_button, 2, 0)
dia_layout.addWidget(labels_dir_label, 2, 1)
# Read only checkbox
read_only_check = QtWidgets.QCheckBox('Read only')
dia_layout.addWidget(read_only_check, 3, 1)
# Read only checkbox
read_only_check = QtWidgets.QCheckBox('Read only')
dia_layout.addWidget(read_only_check, 3, 1)
# DEXTR checkbox
dextr_check = QtWidgets.QCheckBox('Enable DEXTR')
dia_layout.addWidget(dextr_check, 4, 1)
# Buttons
ok_button = QtWidgets.QPushButton('Ok')
ok_button.setIcon(init_dialog.style().standardIcon(
QtWidgets.QStyle.SP_DialogOkButton))
def _on_ok():
init_dialog.accept()
ok_button.clicked.connect(_on_ok)
cancel_button = QtWidgets.QPushButton('Cancel')
cancel_button.setIcon(init_dialog.style().standardIcon(
QtWidgets.QStyle.SP_DialogCancelButton))
def _on_cancel():
init_dialog.reject()
cancel_button.clicked.connect(_on_cancel)
dia_layout.addWidget(ok_button, 5, 0)
dia_layout.addWidget(cancel_button, 5, 1)
# Run the dialog
action = init_dialog.exec()
if action == QtWidgets.QDialog.Accepted:
images_dir = pathlib.Path(images_dir[0])
if labels_dir[0] is not None:
labels_dir = pathlib.Path(labels_dir[0])
else:
labels_dir = None
enable_dextr = dextr_check.checkState()
readonly = read_only_check.checkState()
server = web_server.LabellerServer()
server.start_flask_server()
try:
# If DEXTR is to be made available
if enable_dextr or dextr_weights is not None:
from dextr.model import DextrModel
import torch
# Load the dextr model
device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
if dextr_weights is not None:
dextr_weights = pathlib.Path(dextr_weights).expanduser()
dextr_model = torch.load(dextr_weights,
map_location=device)
else:
dextr_model = DextrModel.pascalvoc_resunet101().to(device)
# Evaluation mode
dextr_model.eval()
# Define a mask prediction function
dextr_fn = lambda image, points: dextr_model.predict(
[image], points[None, :, :])[0] >= 0.5
else:
dextr_fn = None
# Colour schemes
# The user may select different colour schemes for different tasks.
# If you have a lot of classes, it will be difficult to select colours that are easily distinguished
# from one another. For one task e.g. segmentation, design a colour scheme that highlights the different
# classes for that task, while another task e.g. fine-grained classification would use another scheme.
# Each colour scheme is a dictionary containing the following:
# name: symbolic name (Python identifier)
# human_name: human readable name for UI
# These colour schemes are going to split the classes by 'default' (all), natural, and artificial.
# Not really useful, but demonstrates the feature.
colour_schemes = [
dict(name='default', human_name='All'),
dict(name='natural', human_name='Natural'),
dict(name='artificial', human_name='Artifical')
]
# Specify our label classes, organised in groups.
# `LabelClass` parameters are:
# symbolic name (Python identifier)
# human readable name for UI
# and colours by colour scheme, as a dict mapping colour scheme name to RGB value as a list
# The label classes are arranged in groups and will be displayed as such in the UI.
# `LabelClassGroup` parameters are:
# human readable name for UI
# label class (`LabelClass` instance) list
label_classes = [
labelling_tool.LabelClassGroup('Natural', [
labelling_tool.LabelClass(
'tree', 'Trees',
dict(default=[0, 255, 192],
natural=[0, 255, 192],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'lake', 'Lake',
dict(default=[0, 128, 255],
natural=[0, 128, 255],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'flower', 'Flower',
dict(default=[255, 96, 192],
natural=[255, 192, 96],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'leaf', 'Leaf',
dict(default=[65, 255, 0],
natural=[65, 255, 0],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'stem', 'Stem',
dict(default=[128, 64, 0],
natural=[128, 64, 0],
artificial=[128, 128, 128])),
]),
labelling_tool.LabelClassGroup('Artificial', [
labelling_tool.LabelClass(
'building', 'Buildings',
dict(default=[255, 128, 0],
natural=[128, 128, 128],
artificial=[255, 128, 0])),
labelling_tool.LabelClass(
'wall', 'Wall',
dict(default=[0, 128, 255],
natural=[128, 128, 128],
artificial=[0, 128, 255])),
])
]
# Annotation controls
# Labels may also have optional meta-data associated with them
# You could use this for e.g. indicating if an object is fully visible, mostly visible or significantly obscured.
# You could also indicate quality (e.g. blurriness, etc)
# There are four types of annotation. They have some common properties:
# - name: symbolic name (Python identifier)
# - label_text: label text in UI
# Check boxes, radio buttons and popup menus also have:
# - visibility_label_text: [optional] if provided, label visibility can be filtered by this annotation value,
# in which case a drop down will appear in the UI allowing the user to select a filter value
# that will hide/show labels accordingly
# Control types:
# Check box (boolean value):
# `labelling_tool.AnnoControlCheckbox`; only the 3 common parameters listed above
# Radio button (choice from a list):
# `labelling_tool.AnnoControlRadioButtons`; the 3 common parameters listed above and:
# choices: list of `labelling_tool.AnnoControlRadioButtons.choice` that provide:
# value: symbolic value name for choice
# tooltip: extra information for user
# label_on_own_line [optional]: if True, place the label and the buttons on a separate line in the UI
# Popup menu (choice from a grouped list):
# `labelling_tool.AnnoControlPopupMenu`; the 3 common parameters listed above and:
# groups: list of groups `labelling_tool.AnnoControlPopupMenu.group`:
# label_text: label text in UI
# choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# value: symbolic value name for choice
# label_text: choice label text in UI
# tooltip: extra information for user
# Text (free form plain text):
# `labelling_tool.AnnoControlText`; only the 2 common parameters listed above and:
# - multiline: boolean; if True a text area will be used, if False a single line text entry
anno_controls = [
labelling_tool.AnnoControlCheckbox(
'good_quality',
'Good quality',
visibility_label_text='Filter by good quality'),
labelling_tool.AnnoControlRadioButtons(
'visibility',
'Visible',
choices=[
labelling_tool.AnnoControlRadioButtons.choice(
value='full',
label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='mostly',
label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='obscured',
label_text='Obscured',
tooltip='Object is significantly obscured'),
],
label_on_own_line=False,
visibility_label_text='Filter by visibility'),
labelling_tool.AnnoControlPopupMenu(
'material',
'Material',
groups=[
labelling_tool.AnnoControlPopupMenu.group(
label_text='Artifical/buildings',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='concrete',
label_text='Concrete',
tooltip='Concrete objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='plastic',
label_text='Plastic',
tooltip='Plastic objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='asphalt',
label_text='Asphalt',
tooltip='Road, pavement, etc.'),
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Flat natural',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='grass',
label_text='Grass',
tooltip='Grass covered ground'),
labelling_tool.AnnoControlPopupMenu.choice(
value='water',
label_text='Water',
tooltip='Water/lake')
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Vegetation',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='trees',
label_text='Trees',
tooltip='Trees'),
labelling_tool.AnnoControlPopupMenu.choice(
value='shrubbery',
label_text='Shrubs',
tooltip='Shrubs/bushes'),
labelling_tool.AnnoControlPopupMenu.choice(
value='flowers',
label_text='Flowers',
tooltip='Flowers'),
labelling_tool.AnnoControlPopupMenu.choice(
value='ivy',
label_text='Ivy',
tooltip='Ivy')
]),
],
visibility_label_text='Filter by material'),
# labelling_tool.AnnoControlText('comment', 'Comment', multiline=False),
]
image_paths = list(images_dir.glob('*.jpg')) + list(
images_dir.glob('*.png'))
image_paths = [p.absolute() for p in image_paths]
# Create a `PersistentLabelledImage` instance for each image file
labelled_images = labelled_image.LabelledImage.for_image_files(
image_paths, labels_dir=labels_dir, readonly=bool(readonly))
print('Loaded {0} images'.format(len(labelled_images)))
config = web_server.DEFAULT_CONFIG
# Example tasks to appear in checkboxes
tasks = [
dict(name='finished', human_name='[old] finished'),
dict(name='segmentation', human_name='Outlines'),
dict(name='classification', human_name='Classification'),
]
# And a window
win = QtWidgets.QWidget()
win.setWindowTitle('Simple Qt Image Labeller')
# And give it a layout
layout = QtWidgets.QVBoxLayout()
win.setLayout(layout)
# Create the labeller
lbl = controls.QLabellerForLabelledImages(
server=server,
label_classes=label_classes,
labelled_images=labelled_images,
tasks=tasks,
colour_schemes=colour_schemes,
anno_controls=anno_controls,
config=config,
dextr_fn=dextr_fn,
enable_firebug=enable_firebug)
# Create the web engine view
view = QtWebEngineWidgets.QWebEngineView()
# If requested, use a memory-based HTTP cache
# This is very useful if the client-side Javascript code is being developed
# as otherwise chromium's cache will often store old versions of the code that
# will hamper debugging and deveopment
if use_http_memory_cache:
view.page().profile().setHttpCacheType(
QtWebEngineWidgets.QWebEngineProfile.MemoryHttpCache)
# Attach the labeller to the web engine view
lbl.attach_to_web_engine_view(view)
# Add the QWebView to the layout
layout.addWidget(view)
# Show the window and run the app
win.show()
app.exec_()
finally:
print('Stopping server')
server.stop_server()
def run_app(images_pat, labels_dir, readonly, update_label_object_ids,
enable_dextr, dextr_weights):
import os
import glob
import json
import uuid
from image_labelling_tool import labelling_tool
if enable_dextr or dextr_weights is not None:
from dextr.model import DextrModel
import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if dextr_weights is not None:
dextr_weights = os.path.expanduser(dextr_weights)
dextr_model = torch.load(dextr_weights, map_location=device)
else:
dextr_model = DextrModel.pascalvoc_resunet101().to(device)
dextr_model.eval()
dextr_fn = lambda image, points: dextr_model.predict([image], points[
None, :, :])[0] >= 0.5
else:
dextr_fn = None
# Colour schemes
# The user may select different colour schemes for different tasks.
# If you have a lot of classes, it will be difficult to select colours that are easily distinguished
# from one another. For one task e.g. segmentation, design a colour scheme that highlights the different
# classes for that task, while another task e.g. fine-grained classification would use another scheme.
# Each colour scheme is a dictionary containing the following:
# name: symbolic name (Python identifier)
# human_name: human readable name for UI
# These colour schemes are going to split the classes by 'default' (all), natural, and artificial.
# Not really useful, but demonstrates the feature.
colour_schemes = [
dict(name='default', human_name='All'),
dict(name='natural', human_name='Natural'),
dict(name='artificial', human_name='Artifical')
]
# Specify our label classes, organised in groups.
# `LabelClass` parameters are:
# symbolic name (Python identifier)
# human readable name for UI
# and colours by colour scheme, as a dict mapping colour scheme name to RGB value as a list
# The label classes are arranged in groups and will be displayed as such in the UI.
# `LabelClassGroup` parameters are:
# human readable name for UI
# label class (`LabelClass` instance) list
label_classes = [
labelling_tool.LabelClassGroup('Natural', [
labelling_tool.LabelClass(
'tree', 'Trees',
dict(default=[0, 255, 192],
natural=[0, 255, 192],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'lake', 'Lake',
dict(default=[0, 128, 255],
natural=[0, 128, 255],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'flower', 'Flower',
dict(default=[255, 96, 192],
natural=[255, 192, 96],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'leaf', 'Leaf',
dict(default=[65, 255, 0],
natural=[65, 255, 0],
artificial=[128, 128, 128])),
labelling_tool.LabelClass(
'stem', 'Stem',
dict(default=[128, 64, 0],
natural=[128, 64, 0],
artificial=[128, 128, 128])),
]),
labelling_tool.LabelClassGroup('Artificial', [
labelling_tool.LabelClass(
'building', 'Buildings',
dict(default=[255, 128, 0],
natural=[128, 128, 128],
artificial=[255, 128, 0])),
labelling_tool.LabelClass(
'wall', 'Wall',
dict(default=[0, 128, 255],
natural=[128, 128, 128],
artificial=[0, 128, 255])),
])
]
# Annotation controls
# Labels may also have optional meta-data associated with them
# You could use this for e.g. indicating if an object is fully visible, mostly visible or significantly obscured.
# You could also indicate quality (e.g. blurriness, etc)
# There are three types of annotation:
# Check box (boolean value):
# `labelling_tool.AnnoControlCheckbox` parameters:
# name: symbolic name (Python identifier)
# label_text: label text in UI
# Radio button (choice from a list):
# `labelling_tool.AnnoControlRadioButtons` parameters:
# name: symbolic name (Python identifier)
# label_text: label text in UI
# choices: list of `labelling_tool.AnnoControlRadioButtons.choice` that provide:
# value: symbolic value name for choice
# label_text: choice label text in UI
# tooltip: extra information for user
# label_on_own_line [optional]: if True, place the label and the buttons on a separate line in the UI
# Popup menu (choice from a grouped list):
# `labelling_tool.AnnoControlPopupMenu` parameters:
# name: symbolic name (Python identifier)
# label_text: label text in UI
# groups: list of groups `labelling_tool.AnnoControlPopupMenu.group`:
# label_text: group label text in UI
# choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# value: symbolic value name for choice
# label_text: choice label text in UI
# tooltip: extra information for user
anno_controls = [
labelling_tool.AnnoControlCheckbox('good_quality', 'Good quality'),
labelling_tool.AnnoControlRadioButtons(
'visibility',
'Visible',
choices=[
labelling_tool.AnnoControlRadioButtons.choice(
value='full',
label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='mostly',
label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='obscured',
label_text='Obscured',
tooltip='Object is significantly obscured'),
],
label_on_own_line=False),
labelling_tool.AnnoControlPopupMenu(
'material',
'Material',
groups=[
labelling_tool.AnnoControlPopupMenu.group(
label_text='Artifical/buildings',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='concrete',
label_text='Concrete',
tooltip='Concrete objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='plastic',
label_text='Plastic',
tooltip='Plastic objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='asphalt',
label_text='Asphalt',
tooltip='Road, pavement, etc.'),
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Flat natural',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='grass',
label_text='Grass',
tooltip='Grass covered ground'),
labelling_tool.AnnoControlPopupMenu.choice(
value='water',
label_text='Water',
tooltip='Water/lake')
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Vegetation',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='trees', label_text='Trees',
tooltip='Trees'),
labelling_tool.AnnoControlPopupMenu.choice(
value='shrubbery',
label_text='Shrubs',
tooltip='Shrubs/bushes'),
labelling_tool.AnnoControlPopupMenu.choice(
value='flowers',
label_text='Flowers',
tooltip='Flowers'),
labelling_tool.AnnoControlPopupMenu.choice(
value='ivy', label_text='Ivy', tooltip='Ivy')
]),
])
]
if images_pat.strip() == '':
image_paths = glob.glob('images/*.jpg') + glob.glob('images/*.png')
else:
image_paths = glob.glob(images_pat)
# Load in .JPG images from the 'images' directory.
labelled_images = labelling_tool.PersistentLabelledImage.for_files(
image_paths, labels_dir=labels_dir, readonly=readonly)
print('Loaded {0} images'.format(len(labelled_images)))
if update_label_object_ids:
n_updated = 0
for limg in labelled_images:
if os.path.exists(limg.labels_path):
label_js = json.load(open(limg.labels_path, 'r'))
prefix = str(uuid.uuid4())
modified = labelling_tool.ensure_json_object_ids_have_prefix(
label_js, id_prefix=prefix)
if modified:
with open(limg.labels_path, 'w') as f_out:
json.dump(label_js, f_out, indent=3)
n_updated += 1
print('Updated object IDs in {} files'.format(n_updated))
config = {
'tools': {
'imageSelector': True,
'labelClassSelector': True,
'drawPointLabel': False,
'drawBoxLabel': True,
'drawOrientedEllipseLabel': True,
'drawPolyLabel': True,
'compositeLabel': False,
'deleteLabel': True,
'deleteConfig': {
'typePermissions': {
'point': True,
'box': True,
'polygon': True,
'composite': True,
'group': True,
}
}
},
'settings': {
'brushWheelRate':
0.025, # Change rate for brush radius (mouse wheel)
'brushKeyRate': 2.0, # Change rate for brush radius (keyboard)
}
}
tasks = [
dict(name='finished', human_name='[old] finished'),
dict(name='segmentation', human_name='Outlines'),
dict(name='classification', human_name='Classification'),
]
flask_labeller(label_classes,
labelled_images,
tasks=tasks,
colour_schemes=colour_schemes,
anno_controls=anno_controls,
config=config,
dextr_fn=dextr_fn)
def run_app(images_dir, images_pat, labels_dir, readonly, update_label_object_ids,
enable_dextr, dextr_weights):
import pathlib
import json
import uuid
from image_labelling_tool import labelled_image, labelling_tool
if enable_dextr or dextr_weights is not None:
from dextr.model import DextrModel
import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if dextr_weights is not None:
dextr_weights = pathlib.Path(dextr_weights).expanduser()
dextr_model = torch.load(dextr_weights, map_location=device)
else:
dextr_model = DextrModel.pascalvoc_resunet101().to(device)
dextr_model.eval()
dextr_fn = lambda image, points: dextr_model.predict([image], points[None, :, :])[0] >= 0.5
else:
dextr_fn = None
# Colour schemes
# The user may select different colour schemes for different tasks.
# If you have a lot of classes, it will be difficult to select colours that are easily distinguished
# from one another. For one task e.g. segmentation, design a colour scheme that highlights the different
# classes for that task, while another task e.g. fine-grained classification would use another scheme.
# Each colour scheme is a dictionary containing the following:
# name: symbolic name (Python identifier)
# human_name: human readable name for UI
# These colour schemes are going to split the classes by 'default' (all), natural, and artificial.
# Not really useful, but demonstrates the feature.
colour_schemes = [
dict(name='default', human_name='All'),
dict(name='natural', human_name='Natural'),
dict(name='artificial', human_name='Artifical')
]
# Specify our label classes, organised in groups.
# `LabelClass` parameters are:
# symbolic name (Python identifier)
# human readable name for UI
# and colours by colour scheme, as a dict mapping colour scheme name to RGB value as a list
# The label classes are arranged in groups and will be displayed as such in the UI.
# `LabelClassGroup` parameters are:
# human readable name for UI
# label class (`LabelClass` instance) list
label_classes = [
labelling_tool.LabelClassGroup('Natural', [
labelling_tool.LabelClass('tree', 'Trees', dict(default=[0, 255, 192], natural=[0, 255, 192],
artificial=[128, 128, 128])),
labelling_tool.LabelClass('lake', 'Lake', dict(default=[0, 128, 255], natural=[0, 128, 255],
artificial=[128, 128, 128])),
labelling_tool.LabelClass('flower', 'Flower', dict(default=[255, 96, 192], natural=[255, 192, 96],
artificial=[128, 128, 128])),
labelling_tool.LabelClass('leaf', 'Leaf', dict(default=[65, 255, 0], natural=[65, 255, 0],
artificial=[128, 128, 128])),
labelling_tool.LabelClass('stem', 'Stem', dict(default=[128, 64, 0], natural=[128, 64, 0],
artificial=[128, 128, 128])),
]),
labelling_tool.LabelClassGroup('Artificial', [
labelling_tool.LabelClass('building', 'Buildings', dict(default=[255, 128, 0], natural=[128, 128, 128],
artificial=[255, 128, 0])),
labelling_tool.LabelClass('wall', 'Wall', dict(default=[0, 128, 255], natural=[128, 128, 128],
artificial=[0, 128, 255])),
])]
# Annotation controls
# Labels may also have optional meta-data associated with them
# You could use this for e.g. indicating if an object is fully visible, mostly visible or significantly obscured.
# You could also indicate quality (e.g. blurriness, etc)
# There are four types of annotation. They have some common properties:
# - name: symbolic name (Python identifier)
# - label_text: label text in UI
# Check boxes, radio buttons and popup menus also have:
# - visibility_label_text: [optional] if provided, label visibility can be filtered by this annotation value,
# in which case a drop down will appear in the UI allowing the user to select a filter value
# that will hide/show labels accordingly
# Control types:
# Check box (boolean value):
# `labelling_tool.AnnoControlCheckbox`; only the 3 common parameters listed above
# Radio button (choice from a list):
# `labelling_tool.AnnoControlRadioButtons`; the 3 common parameters listed above and:
# - choices: list of `labelling_tool.AnnoControlRadioButtons.choice` that provide:
# - value: symbolic value name for choice
# - tooltip: extra information for user
# - label_on_own_line [optional]: if True, place the label and the buttons on a separate line in the UI
# Popup menu (choice from a grouped list):
# `labelling_tool.AnnoControlPopupMenu`; the 3 common parameters listed above and:
# - groups: list of groups `labelling_tool.AnnoControlPopupMenu.group`:
# - label_text: label text in UI
# - choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# - value: symbolic value name for choice
# - label_text: choice label text in UI
# - tooltip: extra information for user
# Text (free form plain text):
# `labelling_tool.AnnoControlText`; only the 2 common parameters listed above and:
# - multiline: boolean; if True a text area will be used, if False a single line text entry
anno_controls = [
labelling_tool.AnnoControlCheckbox('good_quality', 'Good quality',
visibility_label_text='Filter by good quality'),
labelling_tool.AnnoControlRadioButtons('visibility', 'Visible', choices=[
labelling_tool.AnnoControlRadioButtons.choice(value='full', label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(value='mostly', label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(value='obscured', label_text='Obscured',
tooltip='Object is significantly obscured'),
], label_on_own_line=False, visibility_label_text='Filter by visibility'),
labelling_tool.AnnoControlPopupMenu('material', 'Material', groups=[
labelling_tool.AnnoControlPopupMenu.group(label_text='Artifical/buildings', choices=[
labelling_tool.AnnoControlPopupMenu.choice(value='concrete', label_text='Concrete',
tooltip='Concrete objects'),
labelling_tool.AnnoControlPopupMenu.choice(value='plastic', label_text='Plastic',
tooltip='Plastic objects'),
labelling_tool.AnnoControlPopupMenu.choice(value='asphalt', label_text='Asphalt',
tooltip='Road, pavement, etc.'),
]),
labelling_tool.AnnoControlPopupMenu.group(label_text='Flat natural', choices=[
labelling_tool.AnnoControlPopupMenu.choice(value='grass', label_text='Grass',
tooltip='Grass covered ground'),
labelling_tool.AnnoControlPopupMenu.choice(value='water', label_text='Water', tooltip='Water/lake')]),
labelling_tool.AnnoControlPopupMenu.group(label_text='Vegetation', choices=[
labelling_tool.AnnoControlPopupMenu.choice(value='trees', label_text='Trees', tooltip='Trees'),
labelling_tool.AnnoControlPopupMenu.choice(value='shrubbery', label_text='Shrubs',
tooltip='Shrubs/bushes'),
labelling_tool.AnnoControlPopupMenu.choice(value='flowers', label_text='Flowers',
tooltip='Flowers'),
labelling_tool.AnnoControlPopupMenu.choice(value='ivy', label_text='Ivy', tooltip='Ivy')]),
], visibility_label_text='Filter by material'),
# labelling_tool.AnnoControlText('comment', 'Comment', multiline=False),
]
image_pats = images_pat.split('|')
# Load in .JPG images from the 'images' directory.
labelled_images = labelled_image.LabelledImage.for_directory(
images_dir, image_filename_patterns=image_pats, readonly=readonly)
print('Loaded {0} images'.format(len(labelled_images)))
if update_label_object_ids:
n_updated = 0
for limg in labelled_images:
if limg.labels_store.labels_path.exists():
label_js = json.load(limg.labels_store.labels_path.open('r'))
prefix = str(uuid.uuid4())
modified = labelling_tool.ensure_json_object_ids_have_prefix(
label_js, id_prefix=prefix)
if modified:
with open(limg.labels_store.labels_path, 'w') as f_out:
json.dump(label_js, f_out, indent=3)
n_updated += 1
print('Updated object IDs in {} files'.format(n_updated))
# For documentation of the configuration, please see the comment above `labelling_tool.DEFAULT_CONFIG`
config = labelling_tool.DEFAULT_CONFIG
tasks = [
dict(name='finished', human_name='[old] finished'),
dict(name='segmentation', human_name='Outlines'),
dict(name='classification', human_name='Classification'),
]
flask_labeller(label_classes, labelled_images, tasks=tasks, colour_schemes=colour_schemes,
anno_controls=anno_controls, config=config, dextr_fn=dextr_fn)
def _labeller_window(app,
images_dir: pathlib.Path,
labels_dir: Optional[pathlib.Path],
schema_path: pathlib.Path,
readonly: bool,
enable_dextr: bool,
dextr_weights=Optional[pathlib.Path],
enable_firebug: bool = False,
use_http_memory_cache: bool = False):
from image_labelling_tool import labelled_image, labelling_tool, labelling_schema
server = web_server.LabellerServer()
server.start_flask_server()
try:
# If DEXTR is to be made available
if enable_dextr or dextr_weights is not None:
from dextr.model import DextrModel
import torch
# Load the dextr model
device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
if dextr_weights is not None:
dextr_weights = dextr_weights.expanduser()
dextr_model = torch.load(dextr_weights, map_location=device)
else:
dextr_model = DextrModel.pascalvoc_resunet101().to(device)
# Evaluation mode
dextr_model.eval()
# Define a mask prediction function
dextr_fn = lambda image, points: dextr_model.predict(
[image], points[None, :, :])[0] >= 0.5
else:
dextr_fn = None
schema_store = labelling_schema.FileSchemaStore(schema_path,
readonly=readonly)
# Annotation controls
# Labels may also have optional meta-data associated with them
# You could use this for e.g. indicating if an object is fully visible, mostly visible or significantly obscured.
# You could also indicate quality (e.g. blurriness, etc)
# There are four types of annotation. They have some common properties:
# - name: symbolic name (Python identifier)
# - label_text: label text in UI
# Check boxes, radio buttons and popup menus also have:
# - visibility_label_text: [optional] if provided, label visibility can be filtered by this annotation value,
# in which case a drop down will appear in the UI allowing the user to select a filter value
# that will hide/show labels accordingly
# Control types:
# Check box (boolean value):
# `labelling_tool.AnnoControlCheckbox`; only the 3 common parameters listed above
# Radio button (choice from a list):
# `labelling_tool.AnnoControlRadioButtons`; the 3 common parameters listed above and:
# choices: list of `labelling_tool.AnnoControlRadioButtons.choice` that provide:
# value: symbolic value name for choice
# tooltip: extra information for user
# label_on_own_line [optional]: if True, place the label and the buttons on a separate line in the UI
# Popup menu (choice from a grouped list):
# `labelling_tool.AnnoControlPopupMenu`; the 3 common parameters listed above and:
# groups: list of groups `labelling_tool.AnnoControlPopupMenu.group`:
# label_text: label text in UI
# choices: list of `labelling_tool.AnnoControlPopupMenu.choice` that provide:
# value: symbolic value name for choice
# label_text: choice label text in UI
# tooltip: extra information for user
# Text (free form plain text):
# `labelling_tool.AnnoControlText`; only the 2 common parameters listed above and:
# - multiline: boolean; if True a text area will be used, if False a single line text entry
anno_controls = [
labelling_tool.AnnoControlCheckbox(
'good_quality',
'Good quality',
visibility_label_text='Filter by good quality'),
labelling_tool.AnnoControlRadioButtons(
'visibility',
'Visible',
choices=[
labelling_tool.AnnoControlRadioButtons.choice(
value='full',
label_text='Fully',
tooltip='Object is fully visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='mostly',
label_text='Mostly',
tooltip='Object is mostly visible'),
labelling_tool.AnnoControlRadioButtons.choice(
value='obscured',
label_text='Obscured',
tooltip='Object is significantly obscured'),
],
label_on_own_line=False,
visibility_label_text='Filter by visibility'),
labelling_tool.AnnoControlPopupMenu(
'material',
'Material',
groups=[
labelling_tool.AnnoControlPopupMenu.group(
label_text='Artifical/buildings',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='concrete',
label_text='Concrete',
tooltip='Concrete objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='plastic',
label_text='Plastic',
tooltip='Plastic objects'),
labelling_tool.AnnoControlPopupMenu.choice(
value='asphalt',
label_text='Asphalt',
tooltip='Road, pavement, etc.'),
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Flat natural',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='grass',
label_text='Grass',
tooltip='Grass covered ground'),
labelling_tool.AnnoControlPopupMenu.choice(
value='water',
label_text='Water',
tooltip='Water/lake')
]),
labelling_tool.AnnoControlPopupMenu.group(
label_text='Vegetation',
choices=[
labelling_tool.AnnoControlPopupMenu.choice(
value='trees',
label_text='Trees',
tooltip='Trees'),
labelling_tool.AnnoControlPopupMenu.choice(
value='shrubbery',
label_text='Shrubs',
tooltip='Shrubs/bushes'),
labelling_tool.AnnoControlPopupMenu.choice(
value='flowers',
label_text='Flowers',
tooltip='Flowers'),
labelling_tool.AnnoControlPopupMenu.choice(
value='ivy', label_text='Ivy', tooltip='Ivy')
]),
],
visibility_label_text='Filter by material'),
# labelling_tool.AnnoControlText('comment', 'Comment', multiline=False),
]
image_paths = list(images_dir.glob('*.jpg')) + list(
images_dir.glob('*.png'))
image_paths = [p.absolute() for p in image_paths]
# Create a `PersistentLabelledImage` instance for each image file
labelled_images = labelled_image.LabelledImage.for_image_files(
image_paths, labels_dir=labels_dir, readonly=bool(readonly))
print('Loaded {0} images'.format(len(labelled_images)))
config = web_server.DEFAULT_CONFIG
# Example tasks to appear in checkboxes
tasks = [
dict(name='finished', human_name='[old] finished'),
dict(name='segmentation', human_name='Outlines'),
dict(name='classification', human_name='Classification'),
]
tool = Tool(server=server,
labelled_images=labelled_images,
schema_store=schema_store,
tasks=tasks,
anno_controls=anno_controls,
config=config,
dextr_fn=dextr_fn,
enable_firebug=enable_firebug,
use_http_memory_cache=use_http_memory_cache)
# Show the window and run the app
tool.open_labeller()
app.exec_()
finally:
print('Stopping server')
server.stop_server()