class DeepLabModel(Model):
def __init__(self,config):
super(DeepLabModel, self).__init__(config)
def prepare_input_stream(self):
if self.input_type is 'video':
self._is_videoD = True
self._input_stream = VideoStream(self.config.VIDEO_INPUT,self.config.WIDTH,self.config.HEIGHT).start()
elif self.input_type is 'image':
self._is_imageD = True
self._input_stream = ImageStream(self.config.IMAGE_PATH,self.config.LIMIT_IMAGES).start()
if self.config.WRITE_TIMELINE:
self.prepare_timeliner()
elif self._input_type is 'ros':
self.prepare_ros('deeplab_node')
def prepare_model(self,input_type):
"""
prepares DeepLab model
input_type: must be 'image', 'video', or 'ros'
"""
assert input_type in ['image','video','ros'], "only image, video or ros input possible"
super(DeepLabModel, self).prepare_model()
self.input_type = input_type
# Tracker
if self.config.USE_TRACKER:
self.prepare_tracker()
# Input Stream
self.category_index = None
self.prepare_input_stream()
print("> Building Graph")
with self.detection_graph.as_default():
with tf.Session(graph=self.detection_graph,config=self._tf_config) as self._sess:
return self
def detect(self):
"""
DeepLab Detection function
"""
if not (self.config.USE_TRACKER and self._track):
self.frame = self._input_stream.read()
height,width,_ = self.frame.shape
resize_ratio = 1.0 * 513 / max(self._input_stream.real_width,self._input_stream.real_height)
target_size = (int(resize_ratio * self._input_stream.real_width),int(resize_ratio * self._input_stream.real_height)) #(513, 342)?(513,384)
self.frame = self._visualizer.resize_image(self.frame, target_size)
batch_seg_map = self._sess.run('SemanticPredictions:0',
feed_dict={'ImageTensor:0':
[self._visualizer.convertRGB_image(self.frame)]},
options=self._run_options, run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE:
self._timeliner.write_timeline(self._run_metadata.step_stats,
'{}/timeline_{}.json'.format(
self.config.RESULT_PATH,self.config.DISPLAY_NAME))
seg_map = batch_seg_map[0]
self.boxes = []
self.labels = []
self.ids = []
if self.config.BBOX:
map_labeled = measure.label(seg_map, connectivity=1)
for region in measure.regionprops(map_labeled):
if region.area > self.config.MINAREA:
box = region.bbox
id = seg_map[tuple(region.coords[0])]
label = self.config.LABEL_NAMES[id]
self.boxes.append(box)
self.labels.append(label)
self.ids.append(id)
# deeplab workaround
self.num = len(self.boxes)
self.classes = self.ids
self.scores = self.labels
self.masks = seg_map
# Activate Tracker
if self.config.USE_TRACKER and not self._is_imageD:
self.activate_tracker()
else:
self.run_tracker()
# publish ros
if self._is_rosD:
self._ros_publisher.publish(self.boxes,self.labels,self.masks,self.frame.shape,self.fps.fps_local())
class DeepLabModel(Model):
def __init__(self, config):
super(DeepLabModel, self).__init__(config)
def prepare_model(self, input_type):
"""
prepares DeepLab model
input_type: must be 'image' or 'video'
"""
assert input_type in ['image',
'video'], "only image or video input possible"
super(DeepLabModel, self).prepare_model()
self.input_type = input_type
# fixed input sizes as model needs resize either way
# Input configurations
self.category_index = None
if self.input_type is 'video':
self._input_stream = VideoStream(self.config.VIDEO_INPUT,
self.config.WIDTH,
self.config.HEIGHT).start()
elif self.input_type is 'image':
self._input_stream = ImageStream(self.config.IMAGE_PATH,
self.config.LIMIT_IMAGES).start()
if self.config.WRITE_TIMELINE:
self._run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
self._run_metadata = tf.RunMetadata()
self._timeliner = TimeLiner()
print("> Building Graph")
with self.detection_graph.as_default():
with tf.Session(graph=self.detection_graph,
config=self._tf_config) as self._sess:
return self
def detect(self):
"""
DeepLab Detection function
"""
self.frame = self._input_stream.read()
height, width, _ = self.frame.shape
resize_ratio = 1.0 * 513 / max(self._input_stream.real_width,
self._input_stream.real_height)
target_size = (int(resize_ratio * self._input_stream.real_width),
int(resize_ratio * self._input_stream.real_height)
) #(513, 342)?(513,384)
self.frame = self._visualizer.resize_image(self.frame, target_size)
batch_seg_map = self._sess.run(
'SemanticPredictions:0',
feed_dict={
'ImageTensor:0':
[self._visualizer.convertRGB_image(self.frame)]
},
options=self._run_options,
run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE and self.input_type is 'image':
self._timeliner.write_timeline(
self._run_metadata.step_stats,
'{}/timeline_{}.json'.format(self.config.RESULT_PATH,
self.config.DISPLAY_NAME))
seg_map = batch_seg_map[0]
self.boxes = []
self.labels = []
self.ids = []
if self.config.BBOX:
map_labeled = measure.label(seg_map, connectivity=1)
for region in measure.regionprops(map_labeled):
if region.area > self.config.MINAREA:
box = region.bbox
id = seg_map[tuple(region.coords[0])]
label = self.config.LABEL_NAMES[id]
self.boxes.append(box)
self.labels.append(label)
self.ids.append(id)
# workaround
self.num = len(self.boxes)
self.classes = self.ids
self.scores = self.labels
self.masks = seg_map
class ObjectDetectionModel(Model):
"""
object_detection model class
"""
def __init__(self,config):
super(ObjectDetectionModel, self).__init__(config)
def prepare_input_stream(self):
"""
prepares Input Stream
stream types: 'video','image','ros'
gets called by prepare model
"""
if self.input_type is 'video':
self._is_videoD = True
self._input_stream = VideoStream(self.config.VIDEO_INPUT,self.config.WIDTH,
self.config.HEIGHT).start()
self.stream_height = self._input_stream.real_height
self.stream_width = self._input_stream.real_width
elif self.input_type is 'image':
self._is_imageD = True
self._input_stream = ImageStream(self.config.IMAGE_PATH,self.config.LIMIT_IMAGES,
(self.config.WIDTH,self.config.HEIGHT)).start()
self.stream_height = self.config.HEIGHT
self.stream_width = self.config.WIDTH
elif self.input_type is 'ros':
self.prepare_ros('detection_node')
# Timeliner for image detection
if self.config.WRITE_TIMELINE:
self.prepare_timeliner()
def prepare_model(self,input_type):
"""
prepares Object_Detection model
input_type: must be 'image', 'video', or 'ros'
"""
assert input_type in ['image','video','ros'], "only 'image','video' and 'ros' input possible"
super(ObjectDetectionModel, self).prepare_model()
self.input_type = input_type
# Tracker
if self.config.USE_TRACKER:
self.prepare_tracker()
print("> Building Graph")
with self.detection_graph.as_default():
with tf.Session(graph=self.detection_graph,config=self._tf_config) as self._sess:
# Prepare Input Stream
self.prepare_input_stream()
# Define Input and Ouput tensors
self._tensor_dict = self.get_tensor_dict(['num_detections', 'detection_boxes',
'detection_scores','detection_classes', 'detection_masks'])
self._image_tensor = self.detection_graph.get_tensor_by_name('image_tensor:0')
# Mask Transformations
if 'detection_masks' in self._tensor_dict:
# Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
detection_boxes = tf.squeeze(self._tensor_dict['detection_boxes'], [0])
detection_masks = tf.squeeze(self._tensor_dict['detection_masks'], [0])
real_num_detection = tf.cast(self._tensor_dict['num_detections'][0], tf.int32)
detection_boxes = tf.slice(detection_boxes, [0, 0], [real_num_detection, -1])
detection_masks = tf.slice(detection_masks, [0, 0, 0], [real_num_detection, -1, -1])
detection_masks_reframed = reframe_box_masks_to_image_masks(detection_masks, detection_boxes,self.stream_height,self.stream_width)
detection_masks_reframed = tf.cast(tf.greater(detection_masks_reframed, 0.5), tf.uint8)
self._tensor_dict['detection_masks'] = tf.expand_dims(detection_masks_reframed, 0)
if self.config.SPLIT_MODEL:
self._score_out = self.detection_graph.get_tensor_by_name('{}:0'.format(self.config.SPLIT_NODES[0]))
self._expand_out = self.detection_graph.get_tensor_by_name('{}:0'.format(self.config.SPLIT_NODES[1]))
self._score_in = self.detection_graph.get_tensor_by_name('{}_1:0'.format(self.config.SPLIT_NODES[0]))
self._expand_in = self.detection_graph.get_tensor_by_name('{}_1:0'.format(self.config.SPLIT_NODES[1]))
# Threading
self._gpu_worker = SessionWorker("GPU",self.detection_graph,self._tf_config)
self._cpu_worker = SessionWorker("CPU",self.detection_graph,self._tf_config)
self._gpu_opts = [self._score_out,self._expand_out]
self._cpu_opts = [self._tensor_dict['detection_boxes'],
self._tensor_dict['detection_scores'],
self._tensor_dict['detection_classes'],
self._tensor_dict['num_detections']]
return self
def run_default_sess(self):
"""
runs default session
"""
# default session)
self.frame = self._input_stream.read()
output_dict = self._sess.run(self._tensor_dict,
feed_dict={self._image_tensor:
self._visualizer.expand_and_convertRGB_image(self.frame)},
options=self._run_options, run_metadata=self._run_metadata)
self.num = output_dict['num_detections'][0]
self.classes = output_dict['detection_classes'][0]
self.boxes = output_dict['detection_boxes'][0]
self.scores = output_dict['detection_scores'][0]
if 'detection_masks' in output_dict:
self.masks = output_dict['detection_masks'][0]
def run_thread_sess(self):
"""
runs seperate gpu and cpu session threads
"""
if self._gpu_worker.is_sess_empty():
# put new queue
self.frame = self._input_stream.read()
gpu_feeds = {self._image_tensor: self._visualizer.expand_and_convertRGB_image(self.frame)}
if self.config.VISUALIZE:
gpu_extras = self.frame # for visualization frame
else:
gpu_extras = None
self._gpu_worker.put_sess_queue(self._gpu_opts,gpu_feeds,gpu_extras)
g = self._gpu_worker.get_result_queue()
if g is None:
# gpu thread has no output queue. ok skip, let's check cpu thread.
pass
else:
# gpu thread has output queue.
score,expand,self._frame = g["results"][0],g["results"][1],g["extras"]
if self._cpu_worker.is_sess_empty():
# When cpu thread has no next queue, put new queue.
# else, drop gpu queue.
cpu_feeds = {self._score_in: score, self._expand_in: expand}
cpu_extras = self.frame
self._cpu_worker.put_sess_queue(self._cpu_opts,cpu_feeds,cpu_extras)
c = self._cpu_worker.get_result_queue()
if c is None:
# cpu thread has no output queue. ok, nothing to do. continue
self._wait_thread = True
return # If CPU RESULT has not been set yet, no fps update
else:
self._wait_thread = False
self.boxes,self.scores,self.classes,self.num,self.frame = c["results"][0],c["results"][1],c["results"][2],c["results"][3],c["extras"]
def run_split_sess(self):
"""
runs split session WITHOUT threading
optional: timeline writer
"""
self.frame = self._input_stream.read()
score, expand = self._sess.run(self._gpu_opts,feed_dict={self._image_tensor:
self._visualizer.expand_and_convertRGB_image(self.frame)},
options=self._run_options, run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE:
self.timeliner.write_timeline(self._run_metadata.step_stats,
'{}/timeline_{}_SM1.json'.format(
self.config.RESULT_PATH,self.config.DISPLAY_NAME))
# CPU Session
self.boxes,self.scores,self.classes,self.num = self._sess.run(self._cpu_opts,
feed_dict={self._score_in:score,
self._expand_in: expand},
options=self._run_options,
run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE:
self.timeliner.write_timeline(self._run_metadata.step_stats,
'{}/timeline_{}_SM2.json'.format(
self.config.RESULT_PATH,self.config.DISPLAY_NAME))
def reformat_detection(self):
"""
reformats detection
"""
self.num = int(self.num)
self.boxes = np.squeeze(self.boxes)
self.classes = np.squeeze(self.classes).astype(np.uint8)
self.scores = np.squeeze(self.scores)
def detect(self):
"""
Object_Detection Detection function
optional: multi threading split session, timline writer
"""
if not (self.config.USE_TRACKER and self._track):
if self.config.SPLIT_MODEL:
if self.config.MULTI_THREADING:
self.run_thread_sess()
if self._wait_thread: # checks if thread has output
return
else:
self.run_split_sess()
else:
self.run_default_sess()
if self.config.WRITE_TIMELINE:
self.timeliner.write_timeline(self._run_metadata.step_stats,
'{}/timeline_{}.json'.format(
self.config.RESULT_PATH,self.config.DISPLAY_NAME))
self.reformat_detection()
# Activate Tracker
if self.config.USE_TRACKER and not self._is_imageD:
self.activate_tracker()
# Tracking
else:
self.run_tracker()
# Publish ROS Message
if self._is_rosD:
self._ros_publisher.publish(self.boxes,self.scores,self.classes,self.num,self.category_index,self.frame.shape,self.masks,self.fps.fps_local())
class ObjectDetectionModel(Model):
"""
object_detection model class
"""
def __init__(self, config):
super(ObjectDetectionModel, self).__init__(config)
def prepare_model(self, input_type):
"""
prepares Object_Detection model
input_type: must be 'image' or 'video'
"""
assert input_type in ['image', 'video'
], "only 'image' or 'video' input possible"
super(ObjectDetectionModel, self).prepare_model()
self.input_type = input_type
# Tracker
if self.config.USE_TRACKER:
sys.path.append(os.getcwd() + '/rod/kcf')
import KCF
self._tracker = KCF.kcftracker(False, True, False, False)
self._tracker_counter = 0
self._track = False
print("> Building Graph")
with self.detection_graph.as_default():
with tf.Session(graph=self.detection_graph,
config=self._tf_config) as self._sess:
# Input Configuration
if self.input_type is 'video':
self._input_stream = VideoStream(
self.config.VIDEO_INPUT, self.config.WIDTH,
self.config.HEIGHT).start()
height = self._input_stream.real_height
width = self._input_stream.real_width
elif self.input_type is 'image':
self._input_stream = ImageStream(
self.config.IMAGE_PATH, self.config.LIMIT_IMAGES,
(self.config.WIDTH, self.config.HEIGHT)).start()
height = self.config.HEIGHT
width = self.config.WIDTH
# Timeliner for image detection
if self.config.WRITE_TIMELINE:
self._run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
self._run_metadata = tf.RunMetadata()
self.timeliner = TimeLiner()
# Define Input and Ouput tensors
self._tensor_dict = self.get_tensordict([
'num_detections', 'detection_boxes', 'detection_scores',
'detection_classes', 'detection_masks'
])
self._image_tensor = self.detection_graph.get_tensor_by_name(
'image_tensor:0')
# Mask Transformations
if 'detection_masks' in self._tensor_dict:
# Reframe is required to translate mask from box coordinates to image coordinates and fit the image size.
detection_boxes = tf.squeeze(
self._tensor_dict['detection_boxes'], [0])
detection_masks = tf.squeeze(
self._tensor_dict['detection_masks'], [0])
real_num_detection = tf.cast(
self._tensor_dict['num_detections'][0], tf.int32)
detection_boxes = tf.slice(detection_boxes, [0, 0],
[real_num_detection, -1])
detection_masks = tf.slice(detection_masks, [0, 0, 0],
[real_num_detection, -1, -1])
detection_masks_reframed = reframe_box_masks_to_image_masks(
detection_masks, detection_boxes, height, width)
detection_masks_reframed = tf.cast(
tf.greater(detection_masks_reframed, 0.5), tf.uint8)
self._tensor_dict['detection_masks'] = tf.expand_dims(
detection_masks_reframed, 0)
if self.config.SPLIT_MODEL:
self._score_out = self.detection_graph.get_tensor_by_name(
'{}:0'.format(self.config.SPLIT_NODES[0]))
self._expand_out = self.detection_graph.get_tensor_by_name(
'{}:0'.format(self.config.SPLIT_NODES[1]))
self._score_in = self.detection_graph.get_tensor_by_name(
'{}_1:0'.format(self.config.SPLIT_NODES[0]))
self._expand_in = self.detection_graph.get_tensor_by_name(
'{}_1:0'.format(self.config.SPLIT_NODES[1]))
# Threading
self._gpu_worker = SessionWorker("GPU",
self.detection_graph,
self._tf_config)
self._cpu_worker = SessionWorker("CPU",
self.detection_graph,
self._tf_config)
self._gpu_opts = [self._score_out, self._expand_out]
self._cpu_opts = [
self._tensor_dict['detection_boxes'],
self._tensor_dict['detection_scores'],
self._tensor_dict['detection_classes'],
self._tensor_dict['num_detections']
]
return self
def run_default_sess(self):
"""
runs default session
"""
# default session)
self.frame = self._input_stream.read()
output_dict = self._sess.run(
self._tensor_dict,
feed_dict={
self._image_tensor:
self._visualizer.expand_and_convertRGB_image(self.frame)
},
options=self._run_options,
run_metadata=self._run_metadata)
self.num = output_dict['num_detections'][0]
self.classes = output_dict['detection_classes'][0]
self.boxes = output_dict['detection_boxes'][0]
self.scores = output_dict['detection_scores'][0]
if 'detection_masks' in output_dict:
self.masks = output_dict['detection_masks'][0]
def run_thread_sess(self):
"""
runs seperate gpu and cpu session threads
"""
if self._gpu_worker.is_sess_empty():
# put new queue
self.frame = self._input_stream.read()
gpu_feeds = {
self._image_tensor:
self._visualizer.expand_and_convertRGB_image(self.frame)
}
if self.config.VISUALIZE:
gpu_extras = self.frame # for visualization frame
else:
gpu_extras = None
self._gpu_worker.put_sess_queue(self._gpu_opts, gpu_feeds,
gpu_extras)
g = self._gpu_worker.get_result_queue()
if g is None:
# gpu thread has no output queue. ok skip, let's check cpu thread.
pass
else:
# gpu thread has output queue.
score, expand, self._frame = g["results"][0], g["results"][1], g[
"extras"]
if self._cpu_worker.is_sess_empty():
# When cpu thread has no next queue, put new queue.
# else, drop gpu queue.
cpu_feeds = {self._score_in: score, self._expand_in: expand}
cpu_extras = self.frame
self._cpu_worker.put_sess_queue(self._cpu_opts, cpu_feeds,
cpu_extras)
c = self._cpu_worker.get_result_queue()
if c is None:
# cpu thread has no output queue. ok, nothing to do. continue
self._wait_thread = True
return # If CPU RESULT has not been set yet, no fps update
else:
self._wait_thread = False
self.boxes, self.scores, self.classes, self.num, self.frame = c[
"results"][0], c["results"][1], c["results"][2], c["results"][
3], c["extras"]
def run_split_sess(self):
"""
runs split session WITHOUT threading
optional: timeline writer
"""
self.frame = self._input_stream.read()
score, expand = self._sess.run(
self._gpu_opts,
feed_dict={
self._image_tensor:
self._visualizer.expand_and_convertRGB_image(self.frame)
},
options=self._run_options,
run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE:
self.timeliner.write_timeline(
self._run_metadata.step_stats,
'{}/timeline_{}_SM1.json'.format(self.config.RESULT_PATH,
self.config.DISPLAY_NAME))
# CPU Session
self.boxes, self.scores, self.classes, self.num = self._sess.run(
self._cpu_opts,
feed_dict={
self._score_in: score,
self._expand_in: expand
},
options=self._run_options,
run_metadata=self._run_metadata)
if self.config.WRITE_TIMELINE:
self.timeliner.write_timeline(
self._run_metadata.step_stats,
'{}/timeline_{}_SM2.json'.format(self.config.RESULT_PATH,
self.config.DISPLAY_NAME))
def run_tracker(self):
"""
runs KCF tracker on videoStream frame
!does not work on images, obviously!
"""
self.frame = self._input_stream.read()
if self._first_track:
self._trackers = []
self._tracker_boxes = self.boxes
for box in self.boxes[~np.all(self.boxes == 0, axis=1)]:
self._tracker.init(
conv_detect2track(box, self._input_stream.real_width,
self._input_stream.real_height),
self.tracker_frame)
self._trackers.append(self._tracker)
self._first_track = False
for idx, self._tracker in enumerate(self._trackers):
tracker_box = self._tracker.update(self.frame)
self._tracker_boxes[idx, :] = conv_track2detect(
tracker_box, self._input_stream.real_width,
self._input_stream.real_height)
self._tracker_counter += 1
self.boxes = self._tracker_boxes
# Deactivate Tracker
if self._tracker_counter >= self.config.TRACKER_FRAMES:
self._track = False
self._tracker_counter = 0
def reformat_detection(self):
"""
reformats detection
"""
self.num = int(self.num)
self.boxes = np.squeeze(self.boxes)
self.classes = np.squeeze(self.classes).astype(np.uint8)
self.scores = np.squeeze(self.scores)
def detect(self):
"""
Object_Detection Detection function
optional: multi threading split session, timline writer
"""
if not (self.config.USE_TRACKER and self._track):
if self.config.SPLIT_MODEL:
if self.config.MULTI_THREADING:
self.run_thread_sess()
if self._wait_thread: # checks if thread has output
return
else:
self.run_split_sess()
else:
self.run_default_sess()
if self.config.WRITE_TIMELINE and self.input_type is 'image':
self.timeliner.write_timeline(
self._run_metadata.step_stats,
'{}/timeline_{}.json'.format(self.config.RESULT_PATH,
self.config.DISPLAY_NAME))
self.reformat_detection()
# Activate Tracker
if self.config.USE_TRACKER and self.num <= self.config.NUM_TRACKERS and self.input_type is 'video':
self.tracker_frame = self.frame
self._track = True
self._first_track = True
# Tracking
else:
self.run_tracker()