def __init__(self, data, max_partitions=None, k=None):
"""
:type data: pyspark.RDD
:param data: pyspark RDD (key, k-dim vector like)
:type max_partitions: int
:param max_partitions: maximum number of partition to split into
:type k: int
:param k: dimensionality of the data
Split a given data set into approximately equal sized partition (if max_partitions
is a power of 2 ** k) using binary trees
"""
self.k = int(k) if k is not None else len(data.first()[1])
self.max_partitions = int(
max_partitions) if max_partitions is not None else 4**self.k
current_axis = 0
data.cache()
box = data.aggregate(BoundingBox(k=self.k), lambda total,
(_, v): total.union(BoundingBox(v)),
lambda total, v: total.union(v))
first_partition = data.map(lambda (key, value): ((key, 0), value))
self._create_partitions(first_partition, box)
self.result = data.context.emptyRDD()
for partition in self.partitions.itervalues():
if self.result is None:
self.result = partition
else:
self.result = self.result.union(partition)
def parse_object(osm_object):
object_type = 0
if type(osm_object) == osm.osm.Node:
inserted = False
for bbox in bboxes:
if bbox.__contains__(osm_object):
inserted = True
else:
if bbox.distance_to_osm_object(osm_object) < bbox.get_merge_distance():
bbox.insert_object(osm_object)
inserted = True
if not inserted:
# Some tools, like mapproxy-seed, doesnt want to work with polygons that have all points at the same spot
bbox = BoundingBox([float(osm_object.attribs['lon']) + polygonMinSize,
float(osm_object.attribs['lat']) - polygonMinSize],
[float(osm_object.attribs['lon']),
float(osm_object.attribs['lat'])], mergeDistance, mergePercentage)
bboxes.append(bbox)
object_type = 1
if type(osm_object) == osm.osm.Way:
object_type = 2
if type(osm_object) == osm.osm.Relation:
object_type = 3
return object_type
def mouseReleaseEvent(self, event):
if not self.show_reticle and event.button() == Qt.LeftButton:
return
if event.button() == Qt.LeftButton:
if self.bbox_label is not None:
self.end_pt = event.pos()
rect = self.pt2rect(self.start_pt, self.end_pt)
if rect.width() > 5 and rect.height() > 5:
if self.is_new_tube:
if self.bboxes['current_tube'] is not None:
self.bboxes['other_tubes'].append(
self.bboxes['current_tube'])
self.is_new_tube = False
bbox = BoundingBox.from_qrect(rect, self.bbox_label, 1)
self.bboxes['current_tube'] = bbox
self.bbox_added.emit(self.proj_to_real_img(bbox))
self.update()
elif event.button() == Qt.RightButton:
if not self.mouse_down:
bbox = self.bboxes['current_tube']
if (bbox is not None and
bbox.contain(event.pos().x(), event.pos().y())):
self.bboxes['current_tube'] = None
self.bbox_deleted.emit()
self.update()
self.mouse_down = False
def from_dict(tube_dict):
tube = Tube(tube_dict['id'], tube_dict['label'],
tube_dict['start'], tube_dict['end'])
tube.bboxes = []
if 'bboxes' in tube_dict:
for bbox in tube_dict['bboxes']:
tube.bboxes.append(
BoundingBox(tube.label, bbox['src'], *bbox['bbox']))
return tube
def update(self, frame):
self.track_num += 1
score = super(Tracker, self).update(frame.raw_img)
rect = super(Tracker, self).get_position()
l = int(rect.left())
r = int(rect.right())
t = int(rect.top())
b = int(rect.bottom())
self.bbox = BoundingBox(self.label, 0, l, t, r - l, b - t)
return (self.bbox, score)
def run_images(self, images, cutoff=0.05):
"""
Run the neural net on a batch of input images.
Inputs:
images: batch of pictures as a [b, X, Y, C] numpy array.
values between 0 and 255 (uint8 range)
cutoff: value to clip boxes at.
Returns:
Array of BoundingBox objects representing the found boxes.
"""
k = p.ANCHOR_COUNT
gs = p.GRID_SIZE
batch_size = 1
start_time = time.time()
activations, deltas, gammas, classes, chosen_anchor = \
self.sess.run(self.all_out,
feed_dict={self.inp_batch: images}) # ,
# self.do: 1.0})
print('Took %f seconds!' % (time.time() - start_time))
gammas = np.reshape(gammas, [-1, gs**2 * k])
chosen_anchor = np.reshape(chosen_anchor, [-1, gs**2])
deltas = np.reshape(deltas, [-1, gs**2 * k, 4])
classes = np.reshape(classes, [-1, gs**2 * k, p.OUT_CLASSES])
class_numbers = np.argmax(classes, axis=2)
box_list = []
anchors = u.create_anchors(gs)
for ib in range(batch_size):
boxes = u.delta_to_box(deltas[ib], anchors)
nms_indices = tf.image.non_max_suppression(u.trans_boxes(boxes),
gammas[ib],
5,
iou_threshold=0.0) \
.eval(session=self.sess)
selected_boxes = boxes[nms_indices]
selected_gamma = gammas[ib, nms_indices]
selected_class = class_numbers[ib, nms_indices]
selected_class_scores = classes[ib, nms_indices]
for i, box in enumerate(selected_boxes):
sm_scores = u.softmax(selected_class_scores[i])
conf = selected_gamma[i] * sm_scores[selected_class[i]]
if conf > cutoff:
print(conf)
box_list.append(
BoundingBox(u.trans_boxes(box), conf,
selected_class[i]))
return box_list
def display(self, pixmap):
self.scale_ratio = max(pixmap.width() / self.width(),
pixmap.height() / self.height())
scaled_pixmap = pixmap.scaled(self.width() - 2, self.height() - 2,
Qt.KeepAspectRatio)
x = int((self.width() - scaled_pixmap.width()) / 2)
y = int((self.height() - scaled_pixmap.height()) / 2)
self.img_region = BoundingBox.from_qrect(
QRect(QPoint(x, y), scaled_pixmap.size()))
self.setPixmap(scaled_pixmap)
self.update()
def __build_windows(self, outputs: np.ndarray):
windows = []
for size, values in zip(self.window_size, outputs):
values = np.squeeze(values)
rows, cols = values.shape
for i in range(rows):
for j in range(cols):
if values[i, j] > self.threshold:
windows.append((
BoundingBox(j * size * self.window_step,
i * size * self.window_step,
size * INPUT_WIDTH,
size * INPUT_HEIGHT),
size,
values[i, j]
))
return windows
def find_dogs(self, image_path: pathlib.Path) -> None:
"""
Search for dogs in the specified image using a convolutional implementation of sliding windows.
:param image_path: Path to the image to evaluate.
"""
# Loads image
self.image = Image.open(image_path)
# Pre-computes approximately the total number of windows
total_windows = sum([
math.ceil((self.image.width - int(INPUT_WIDTH * size) + 1) /
self.window_step) * math.ceil(
(self.image.height - int(INPUT_HEIGHT * size) + 1) /
self.window_step) for size in self.window_size
])
# Goes through every window detecting if there is a dog in the window
self.__setup_progress(total_windows)
windows = []
for size in self.window_size:
resized_image = np.array(
self.image.resize((int(
self.image.width / size), int(self.image.height / size)),
resample=1))
height, width, _ = resized_image.shape
for i in range(0, height - INPUT_HEIGHT + 1,
max(1, int(self.window_step / size))):
for j in range(0, width - INPUT_WIDTH + 1,
max(1, int(self.window_step / size))):
cropped_image = resized_image[i:i + INPUT_HEIGHT,
j:j + INPUT_WIDTH, :]
probability = self.model.predict(
self.preprocess_fnc(cropped_image))[0][0]
# Window is consider only when the probability is higher than some threshold
if probability >= self.threshold:
windows.append((BoundingBox(j * size, i * size,
size * INPUT_WIDTH,
size * INPUT_HEIGHT), size,
probability))
self.__update_progress()
# Reduces the number of windows detected
self.dogs = self.__filter_windows(windows)
# Plots final windows
self.__plot_windows(self.dogs)
def run(self):
self.centers = []
bboxes = []
self.connection.write(struct.pack('<c', b'd'))
self.connection.flush() #Send b'c' to NNServer to request data
box_count = struct.unpack('<L',
self.connection.read(
struct.calcsize('<L')))[0]
self.data_stream.seek(0)
for i in range(box_count):
data = struct.unpack(
'<ffffff', self.connection.read(struct.calcsize('<ffffff')))
coords = (data[0], data[1], data[2], data[3])
classification = data[4]
confidence = data[5]
bboxes.append(BoundingBox(coords, confidence, classification))
self.data_stream.seek(0)
self.centers = self.box_to_center(bboxes)
'''except:
def test_start_inside_box():
bbox = BB(P(0, 0, 0), P(1, 1, 1))
assert bbox.is_hit(R(P(0.5, 0.5, 0.5), V(1, 1, 1)))
def test_ray_rides_edge_of_bbox():
bbox = BB(P(0, 0, 0), P(1, 1, 1))
assert bbox.is_hit(R(P(0, 0, 0), V(1, 0, 0)))
def test_reject_ray_starting_on_edge_and_moving_away():
# t0 < epsilon
bbox = BB(P(0, 0, 0), P(1, 1, 1))
assert not bbox.is_hit(R(P(0, 0, 0), V(-1, -1, -1)))
def track(self, frame):
frame_rect = BoundingBox(None, 0, 0, 0, self.video.width,
self.video.height)
bbox, score = self.tracker.update(frame)
bbox = bbox.intersected(frame_rect)
return bbox
def test_ray_parallel_to_bbox():
bbox = BB(P(1, 1, 1), P(2, 2, 2))
assert not bbox.is_hit(R(P(0, 0, 0), V(1, 0, 0)))
assert not bbox.is_hit(R(P(0, 0, 0), V(-1, 0, 0)))
def test_bbox_basic_hit():
bbox = BB(P(1, 1, 1), P(2, 2, 2))
assert bbox.is_hit(R(P(0, 0, 0), V(1, 1, 1)))
def test_bbox_basic_miss():
bbox = BB(P(1, 1, 1), P(2, 2, 2))
assert not bbox.is_hit(R(P(0, 0, 0), V(-1, 1, 1)))
assert not bbox.is_hit(R(P(0, 0, 0), V(1, -1, 1)))
assert not bbox.is_hit(R(P(0, 0, 0), V(1, 1, -1)))
def test_hit_box_behind_ray():
bbox = BB(P(1, 1, 1), P(2, 2, 2))
assert not bbox.is_hit(R(P(0, 0, 0), V(-1, -1, -1)))
