def pr_stats(run, image_dir, label_db, connected_components_threshold):
# TODO: a bunch of this can go back into one off init in a class
_train_opts, model = m.restore_model(run)
label_db = LabelDB(label_db_file=label_db)
set_comparison = u.SetComparison()
# use 4 images for debug
debug_imgs = []
for idx, filename in enumerate(sorted(os.listdir(image_dir))):
# load next image
# TODO: this block used in various places, refactor
img = np.array(Image.open(image_dir + "/" +
filename)) # uint8 0->255 (H, W)
img = img.astype(np.float32)
img = (img / 127.5) - 1.0 # -1.0 -> 1.0 # see data.py
# run through model
prediction = expit(model.predict(np.expand_dims(img, 0))[0])
if len(debug_imgs) < 4:
debug_imgs.append(u.side_by_side(rgb=img, bitmap=prediction))
# calc [(x,y), ...] centroids
predicted_centroids = u.centroids_of_connected_components(
prediction, rescale=2.0, threshold=connected_components_threshold)
# compare to true labels
true_centroids = label_db.get_bugs(filename)
true_centroids = [(y, x) for (x, y) in true_centroids] # sigh...
tp, fn, fp = set_comparison.compare_sets(true_centroids,
predicted_centroids)
precision, recall, f1 = set_comparison.precision_recall_f1()
return {
"debug_imgs": debug_imgs,
"precision": precision,
"recall": recall,
"f1": f1
}
class LabelUI(QGraphicsView):
"""PyQt image viewer adapted from
https://github.com/marcel-goldschen-ohm/PyQtImageViewer/blob/master/QtImageViewer.py
"""
# Create signals for mouse events. Note that mouse button signals
# emit (x, y) coordinates but image matrices are indexed (y, x).
leftMouseButtonPressed = pyqtSignal(float, float)
rightMouseButtonPressed = pyqtSignal(float, float)
leftMouseButtonReleased = pyqtSignal(float, float)
rightMouseButtonReleased = pyqtSignal(float, float)
leftMouseButtonDoubleClicked = pyqtSignal(float, float)
rightMouseButtonDoubleClicked = pyqtSignal(float, float)
def __init__(self, label_db_filename, img_dir):
QGraphicsView.__init__(self)
self.setWindowTitle(label_db_filename)
if img_dir is None:
img_dir = str(
QFileDialog.getExistingDirectory(self,
'Select image directory'))
if not os.path.exists(img_dir):
raise RuntimeError(f'Provided directory {img_dir} does not exist')
self.img_dir = img_dir
files_list = []
# Walk through directory tree, get all files
for dir_path, dir_names, filenames in os.walk(img_dir):
files_list += [os.path.join(dir_path, f) for f in filenames]
files_list = sorted(files_list)
files_list = list(
filter(
lambda x: x.lower().endswith(('.png', '.jpg', '.jpeg', '.tiff',
'.bmp', '.gif', '.cr2')),
files_list))
self.files = files_list
if len(self.files) == 0:
raise RuntimeError(
f'Unable to find any image files in provided directory {img_dir}'
)
# Label db
self.label_db = LabelDB(label_db_filename)
self.label_db.create_if_required()
# A lookup table from bug x,y to any labels that have been added
self.x_y_to_labels = {} # { (x, y): Label, ... }
# Flag to denote if bugs are being displayed or not.
# While not displayed, we lock down all image navigation
self.display_labels = True
# Main review loop
self.file_idx = 0
# Image is displayed as a QPixmap in a QGraphicsScene attached to this QGraphicsView
self.scene = QGraphicsScene()
self.setScene(self.scene)
# Store a local handle to the scene's current image pixmap
self._pixmapHandle = None
# Scale image to fit inside viewport, preserving aspect ratio
self.aspectRatioMode = Qt.KeepAspectRatio
# Shows a scroll bar only when zoomed
self.setHorizontalScrollBarPolicy(Qt.ScrollBarAsNeeded)
self.setVerticalScrollBarPolicy(Qt.ScrollBarAsNeeded)
# Stack of QRectF zoom boxes in scene coordinates
self.zoomStack = []
# Initialize some other variables used occasionally
self.tmp_x_y = []
self.click_start_pos = QPoint(0, 0)
self._t_key_pressed = False
self._started_tickmark_click = False
self.complete = False
self.display_image()
self.show()
self.setWindowState(Qt.WindowMaximized)
def update_title(self):
name = os.path.basename(self.files[self.file_idx])
num_bugs = 0
num_tickmarks = 0
num_tickmark_numbers = 0
for label in self.x_y_to_labels.values():
if isinstance(label, Bug):
num_bugs += 1
elif isinstance(label, Tickmark):
num_tickmarks += 1
elif isinstance(label, TickmarkNumber):
num_tickmark_numbers += 1
title = f'{name} ({self.file_idx + 1} of {len(self.files)}): '
title += f'{num_bugs} bug'
if num_bugs != 1:
title += 's'
title += f', {num_tickmarks} tickmark'
if num_tickmarks != 1:
title += 's'
title += f', {num_tickmark_numbers} tickmark number'
if num_tickmark_numbers != 1:
title += 's'
if self.complete:
title += ' [COMPLETE]'
self.setWindowTitle(title)
def has_image(self):
""" Returns whether or not the scene contains an image pixmap.
"""
return self._pixmapHandle is not None
def clear_image(self):
""" Removes the current image pixmap from the scene if it exists.
"""
if self.has_image():
self.scene.removeItem(self._pixmapHandle)
self._pixmapHandle = None
def pixmap(self):
""" Returns the scene's current image pixmap as a QPixmap, or else None if no image exists.
:rtype: QPixmap | None
"""
if self.has_image():
return self._pixmapHandle.pixmap()
return None
def set_image(self, image):
""" Set the scene's current image pixmap to the input QImage or QPixmap.
Raises a RuntimeError if the input image has type other than QImage or QPixmap.
"""
if type(image) is QPixmap:
pixmap = image
elif isinstance(image, QImage):
pixmap = QPixmap.fromImage(image)
else:
raise RuntimeError(
'ImageViewer.setImage: Argument must be a QImage or QPixmap.')
if self.has_image():
self._pixmapHandle.setPixmap(pixmap)
else:
self._pixmapHandle = self.scene.addPixmap(pixmap)
self.setSceneRect(QRectF(
pixmap.rect())) # Set scene size to image size
self.update_viewer()
def update_viewer(self):
""" Show current zoom (if showing entire image, apply current aspect ratio mode).
"""
if not self.has_image():
return
if len(self.zoomStack) and self.sceneRect().contains(
self.zoomStack[-1]):
self.fitInView(
self.zoomStack[-1],
Qt.IgnoreAspectRatio) # Show zoomed rect (ignore aspect ratio)
else:
# Clear the zoom stack (in case we got here because of an invalid zoom)
self.zoomStack = []
# Show entire image (use current aspect ratio mode)
self.fitInView(self.sceneRect(), self.aspectRatioMode)
def resizeEvent(self, event):
""" Maintain current zoom on resize.
"""
self.update_viewer()
def mousePressEvent(self, event):
""" Start mouse pan or zoom mode.
"""
scene_pos = self.mapToScene(event.pos())
self.click_start_pos = event.pos()
if event.button() == Qt.LeftButton:
if self._t_key_pressed:
self._started_tickmark_click = True
self.setDragMode(QGraphicsView.RubberBandDrag)
else:
self.setDragMode(QGraphicsView.ScrollHandDrag)
self.leftMouseButtonPressed.emit(scene_pos.x(), scene_pos.y())
elif event.button() == Qt.RightButton:
self.setDragMode(QGraphicsView.RubberBandDrag)
self.rightMouseButtonPressed.emit(scene_pos.x(), scene_pos.y())
QGraphicsView.mousePressEvent(self, event)
def mouseReleaseEvent(self, event):
""" Stop mouse pan or zoom mode (apply zoom if valid).
"""
QGraphicsView.mouseReleaseEvent(self, event)
scene_pos = self.mapToScene(event.pos())
movement_vector = event.pos() - self.click_start_pos
click_distance = math.sqrt(movement_vector.x()**2 +
movement_vector.y()**2)
if event.button() == Qt.LeftButton:
if self._started_tickmark_click:
if click_distance < 1:
self.add_tickmark_event(event)
else:
view_bbox = self.zoomStack[-1] if len(
self.zoomStack) else self.sceneRect()
selection_bbox = self.scene.selectionArea().boundingRect(
).intersected(view_bbox)
self.scene.setSelectionArea(
QPainterPath()) # Clear current selection area.
if selection_bbox.isValid() and (selection_bbox !=
view_bbox):
self.add_tickmark_number_event(selection_bbox)
# A little hacky, assume someone releases T key to type the numbers.
# If they keep it pressed down they'll be switched back to bug mode
# and will have to press it again.
self._t_key_pressed = False
else:
if click_distance < 1:
self.add_bug_event(event)
self.setDragMode(QGraphicsView.NoDrag)
self.leftMouseButtonReleased.emit(scene_pos.x(), scene_pos.y())
elif event.button() == Qt.RightButton:
view_bbox = self.zoomStack[-1] if len(
self.zoomStack) else self.sceneRect()
selection_bbox = self.scene.selectionArea().boundingRect(
).intersected(view_bbox)
self.scene.setSelectionArea(
QPainterPath()) # Clear current selection area.
if selection_bbox.isValid() and (selection_bbox != view_bbox):
self.zoomStack.append(selection_bbox)
self.update_viewer()
self.setDragMode(QGraphicsView.NoDrag)
if click_distance < 1:
self.remove_closest_label_event(event)
self.rightMouseButtonReleased.emit(scene_pos.x(), scene_pos.y())
self._started_tickmark_click = False
self.update_title()
def keyPressEvent(self, event):
if event.key() == Qt.Key_Right:
self.display_next_image()
elif event.key() == Qt.Key_Left:
self.display_previous_image()
elif event.key() == Qt.Key_Up:
self.toggle_bugs()
elif event.key() == Qt.Key_N:
self.display_next_incomplete_image()
elif event.key() == Qt.Key_Q:
self.exit_program()
elif event.key() == Qt.Key_Escape:
self.zoomStack = [] # Clear zoom stack.
self.update_viewer()
elif event.key() == Qt.Key_T:
self._t_key_pressed = True
elif event.key() == Qt.Key_C:
self.complete = False if self.complete else True
self.update_title()
def exit_program(self):
self._flush_pending_x_y_to_boxes()
QApplication.instance().quit()
def keyReleaseEvent(self, event):
if event.key() == Qt.Key_T:
self._t_key_pressed = False
def display_image(self):
# Open image
img_name = self.files[self.file_idx]
img_path = os.path.join(self.img_dir, img_name)
# If this is a raw file it gets special treatment
if img_path.lower().endswith('.cr2'):
# Read raw file
raw = rawpy.imread(img_path)
# Convert to PIL Image
img = Image.fromarray(raw.postprocess())
else:
img = Image.open(img_path)
# For some reason RGB images do not like to display in the interface.
# RGBA seems to work
img = img.convert('RGBA')
# Convert to QImage
img = ImageQt(img)
self.set_image(img)
# Look up any existing labels in DB for this image and add them
self.update_state_from_db(img_name)
def update_state_from_db(self, img_name):
existing_bugs = self.label_db.get_bugs(img_name)
for x, y in existing_bugs:
self.add_bug_at(x, y)
existing_tickmarks = self.label_db.get_tickmarks(img_name)
for x, y in existing_tickmarks:
self.add_tickmark_at(x, y)
existing_tickmark_numbers = self.label_db.get_tickmark_numbers(
img_name)
for x, y, w, h, val in existing_tickmark_numbers:
self.add_tickmark_number_at(x, y, w, h, val)
complete = self.label_db.get_complete(img_name)
self.complete = complete
self.update_title()
def display_next_image(self):
if not self.display_labels:
print('ignore move to next image; labels not displayed')
return
self._flush_pending_x_y_to_boxes()
self.file_idx += 1
if self.file_idx == len(self.files):
print("Can't move to image past last image.")
self.file_idx = len(self.files) - 1
self.display_image()
def display_previous_image(self):
if not self.display_labels:
print('ignore move to previous image; labels not displayed')
return
self._flush_pending_x_y_to_boxes()
self.file_idx -= 1
if self.file_idx < 0:
print("Can't move to image previous to first image.")
self.file_idx = 0
self.display_image()
def display_next_incomplete_image(self):
self._flush_pending_x_y_to_boxes()
while True:
self.file_idx += 1
if self.file_idx == len(self.files):
print("Can't move to image past last image.")
self.file_idx = len(self.files) - 1
break
if not self.label_db.get_complete(self.files[self.file_idx]):
break
self.display_image()
def add_bug_at(self, x, y):
size = self.scene.width() // 300
rectangle_id = self.scene.addRect(x - size // 2, y - size // 2, size,
size, QPen(Qt.black), QBrush(Qt.red))
self.x_y_to_labels[(x, y)] = Bug(x, y, rectangle_id)
self.update_title()
def add_bug_event(self, e):
scene_pos = self.mapToScene(e.pos())
if not self.display_labels:
print('ignore add bug; labels not displayed')
return
self.add_bug_at(scene_pos.x(), scene_pos.y())
def add_tickmark_at(self, x, y):
size = self.scene.width() // 300
rectangle_id = self.scene.addRect(x - size // 2, y - size // 2,
size, size, QPen(Qt.black),
QBrush(Qt.blue))
self.x_y_to_labels[(x, y)] = Tickmark(x, y, rectangle_id)
self.update_title()
def add_tickmark_event(self, e):
scene_pos = self.mapToScene(e.pos())
if not self.display_labels:
print('ignore add tickmark; labels not displayed')
return
self.add_tickmark_at(scene_pos.x(), scene_pos.y())
def add_tickmark_number_at(self, x, y, width, height, val):
rectangle_id = self.scene.addRect(
x, y, width, height, QPen(Qt.blue,
self.scene.width() // 300))
font = QFont()
font_pixel_size = 1
font.setPixelSize(font_pixel_size)
while QFontMetrics(font).boundingRect(str(val)).width() < width \
and QFontMetrics(font).boundingRect(str(val)).height() < height:
font_pixel_size += 1
font.setPixelSize(font_pixel_size)
font_pixel_size -= 1
number_canvas_id = self.scene.addText(str(val), font)
number_canvas_id.setDefaultTextColor(Qt.blue)
number_canvas_id.setPos(QPoint(int(x), int(y)))
self.x_y_to_labels[(x, y)] = TickmarkNumber(x, y, rectangle_id, width,
height, val,
number_canvas_id)
def add_tickmark_number_event(self, box):
if not self.display_labels:
print('ignore add tickmark; labels not displayed')
return
val, _ = QInputDialog.getInt(self, 'Input', 'Enter tickmark value:')
self.add_tickmark_number_at(box.x(), box.y(), box.width(),
box.height(), val)
def _flush_pending_x_y_to_boxes(self):
"""Write labels to database and remove them from canvas
"""
img_name = self.files[self.file_idx]
# Write to database
self.label_db.set_labels(img_name, self.x_y_to_labels.values())
# Remove from canvas
for label in self.x_y_to_labels.values():
self.remove_label(label)
self.x_y_to_labels.clear()
self.label_db.set_complete(img_name, self.complete)
def toggle_bugs(self):
# if self.display_labels:
# # store x,y s in tmp list and delete all rectangles from canvas
# self.tmp_x_y = []
# for (x, y), rectangle_id in self.x_y_to_labels.items():
# self.remove_label(rectangle_id)
# self.tmp_x_y.append((x, y))
# self.x_y_to_labels = {}
# self.display_labels = False
# else: # labels not displayed
# # restore all temp stored bugs
# for x, y in self.tmp_x_y:
# self.add_bug_at(x, y)
# self.display_labels = True
self.display_labels = not self.display_labels
for item in self.scene.items():
if not isinstance(item, QGraphicsPixmapItem):
item.setVisible(self.display_labels)
def remove_label(self, label):
canvas_id = label.canvas_id
self.scene.removeItem(canvas_id)
if isinstance(label, TickmarkNumber):
number_canvas_id = label.number_canvas_id
self.scene.removeItem(number_canvas_id)
def remove_closest_label_event(self, e):
scene_pos = self.mapToScene(e.pos())
if not self.display_labels:
print('ignore remove label; labels not displayed')
return
if len(self.x_y_to_labels) == 0:
return
closest_point = None
closest_sqr_distance = 0.0
for x, y in self.x_y_to_labels.keys():
sqr_distance = (scene_pos.x() - x)**2 + (scene_pos.y() - y)**2
if sqr_distance < closest_sqr_distance or closest_point is None:
closest_point = (x, y)
closest_sqr_distance = sqr_distance
self.remove_label(self.x_y_to_labels.pop(closest_point))
self.update_title()
for filename in filenames:
original_filename = filename
filename = bnn_util.get_path_relative_to_drive(filename)
drive_base_path = os.path.expanduser(
'~/data/srpa226-drive/Sharing/202012 Paul/')
filename = os.path.join(drive_base_path, filename)
if not os.path.exists(filename):
print(f'File not found, skipping: {filename}')
else:
img = Image.open(filename)
width, height = img.size
if not label_db.get_complete(original_filename):
print(f'Image labeling not complete, skipping: {filename}')
else:
bitmap = bnn_util.xys_to_bitmap(
xys=label_db.get_bugs(original_filename),
height=height,
width=width,
rescale=opts.label_rescale)
single_channel_img = bnn_util.bitmap_to_single_channel_pil_image(
bitmap)
print(f'Processing {filename}')
img_new_filename = os.path.join(opts.image_output_dir,
os.path.basename(filename))
bitmap_filename = os.path.basename(
os.path.splitext(filename)[0] + '_train_bitmap_bugs.png')
bitmap_filename = os.path.join(opts.label_output_dir,
bitmap_filename)
img.save(img_new_filename)
single_channel_img.save(bitmap_filename)