def run(self, item, progress=None):
assert isinstance(item, dl.Item)
logger.info('Downloading video')
video_filepath = item.download()
try:
logger.info('Running dection on video')
annotations_dict = detect_video(graph=self.graph,
yolo=self.yolo,
encoder=self.encoder,
video_filepath=video_filepath,
mark_on_video=False,
show=False)
logger.info('Building Dataloop annotations')
annotations_builder = item.annotations.builder()
for i_frame, annotations in annotations_dict.items():
for annotation in annotations:
annotations_builder.add(
annotation_definition=dl.Box(
top=annotation['top'],
left=annotation['left'],
bottom=annotation['bottom'],
right=annotation['right'],
label=annotation['label']),
object_id=annotation['object_id'],
frame_num=i_frame,
end_frame_num=i_frame,
)
logger.info('Uploading annotations...')
item.annotations.upload(annotations_builder)
logger.info(
'Done! detection and annotations finished successfully')
finally:
if os.path.isfile(video_filepath):
os.remove(video_filepath)
def predict_item(self,
item,
checkpoint_path=None,
with_upload=True,
model_name='object_detection'):
filepath = item.download()
results_path = self.predict_single_image(
image_path=filepath, checkpoint_path=checkpoint_path)
if with_upload:
with open(results_path) as fg:
results = fg.readlines()
builder = item.annotations.builder()
for result in results:
result_ls = result.split(' ')
builder.add(dl.Box(left=int(result_ls[2]),
top=int(result_ls[3]),
right=int(result_ls[4]),
bottom=int(result_ls[5]),
label=result_ls[0]),
model_info={
'confidence': result_ls[1],
'name': model_name
})
item.annotations.upload(builder)
dirname = os.path.dirname(filepath)
return dirname
def main():
"""
This is an example how to upload files and annotations to Dataloop platform.
Image folder contains the images to upload.
Annotations folder contains json file of the annotations. Same name as the image.
We read the images one by one and create the Dataloop annotations using the annotation builder.
Finally, we upload both the image and the matching annotations
:return:
"""
import json
import os
import dtlpy as dl
# Get project and dataset
project = dl.projects.get(project_name='Yachts')
dataset = project.datasets.get(dataset_name='Open Seas')
images_folder = '/home/local/images'
annotations_folder = '/home/local/annotations'
for img_filename in os.listdir(images_folder):
# get the matching annotations json
_, ext = os.path.splitext(img_filename)
ann_filename = os.path.join(annotations_folder,
img_filename.replace(ext, '.json'))
img_filename = os.path.join(images_folder, img_filename)
# Upload or get annotations from platform (if already exists)
item = dataset.items.upload(local_path=img_filename,
remote_path='/in_storm',
overwrite=False)
assert isinstance(item, dl.Item)
# read annotations from file
with open(ann_filename, 'r') as f:
annotations = json.load(f)
# create a Builder instance and add all annotations to it
builder = item.annotations.builder()
for annotation in annotations:
# line format if 4 points of bbox
# this is where you need to update according to your annotation format
label_id = annotation['label']
left = annotation['left']
top = annotation['top']
right = annotation['right']
bottom = annotation['bottom']
builder.add(annotation_definition=dl.Box(top=top,
left=left,
bottom=bottom,
right=right,
label=str(label_id)))
# upload annotations
item.annotations.upload(builder)
def my_converter(annotation):
"""
:param annotation: custom annotation format
:type annotation: dict
:return: dataloop Annotation object
"""
annotations = dl.Annotation.new(annotation_definition=dl.Box(top=annotation['top'],
bottom=annotation['bottom'],
left=annotation['left'],
right=annotation['right'],
label=annotation['label']))
return annotation
def main():
import dtlpy as dl
import numpy as np
import matplotlib.pyplot as plt
# Get project and dataset
project = dl.projects.get(project_name='Food')
dataset = project.datasets.get(dataset_name='BeansDataset')
# get item from platform
item = dataset.items.get(filepath='/image.jpg')
# Create a builder instance
builder = item.annotations.builder()
# add annotations of type box and label person
builder.add(annotation_definition=dl.Box(
top=10, left=10, bottom=100, right=100, label='black_bean'))
# add annotations of type point with attribute
builder.add(annotation_definition=dl.Point(x=80, y=80, label='pea'),
attribute=['large'])
# add annotations of type polygon
builder.add(annotation_definition=dl.Polyline(
geo=[[80, 40], [100, 120], [110, 130]], label='beans_can'))
# create a mask
mask = np.zeros(shape=(item.height, item.width), dtype=np.uint8)
# mark some part in the middle
mask[50:100, 200:250] = 1
# add annotations of type segmentation
builder.add(
annotation_definition=dl.Segmentation(geo=mask, label='tomato_sauce'))
# plot the all of the annotations you created
plt.figure()
plt.imshow(builder.show())
for annotation in builder:
plt.figure()
plt.imshow(annotation.show())
plt.title(annotation.label)
# upload annotations to the item
item.annotations.upload(builder)
def main():
import dtlpy as dl
import matplotlib.pyplot as plt
# Get project and dataset
project = dl.projects.get(project_name='Food')
dataset = project.datasets.get(dataset_name='BeansDataset')
item = dataset.items.get(filepath='/flying boxes.mp4')
############################
# using annotation builder #
############################
# create annotation builder
builder = item.annotations.builder()
for i_frame in range(100):
# go over 100 frame
for i_detection in range(10):
# for each frame we have 10 different detections (location is just for the example)
builder.add(
annotation_definition=dl.Box(top=2 * i_frame,
left=2 * i_detection,
bottom=2 * i_frame + 10,
right=2 * i_detection + 100,
label="moving box"),
frame_num=i_frame, # set the frame for the annotation
object_id=i_detection + 1
# need to input the element id to create the connection between frames
)
# starting from frame 50 add another 10 new annotations of a moving point
if i_frame > 50:
builder.add(annotation_definition=dl.Point(
x=2 * i_frame, y=2 * i_detection, label="moving point"),
frame_num=i_frame,
object_id=20 + (i_detection + 1))
# get frame annotations
frame_annotations = builder.get_frame(frame_num=55)
# Plot the annotations in frame 55 of the created annotations
plt.figure()
plt.imshow(frame_annotations.show())
# plot each annotations separately
for annotation in frame_annotations:
plt.figure()
plt.imshow(annotation.show())
plt.title(annotation.label)
# Add the annotations to platform
item.annotations.upload(builder)
#####################
# single annotation #
#####################
# create one annotations for a video (without using tracker and object ids)
annotation = dl.Annotation.new(item=item)
for i_frame in range(100):
# go over 100 frame
annotation.add_frame(
annotation_definition=dl.Box(
top=2 * i_frame,
left=2 * (i_frame + 10),
bottom=2 * (i_frame + 50),
right=2 * (i_frame + 100),
label="moving box",
),
frame_num=i_frame, # set the frame for the annotation
)
# upload to platform
annotation.upload()
##############################################
# show annotation state in a specified frame #
##############################################
# Get from platform
annotations = item.annotations.list()
# Plot the annotations in frame 55 of the created annotations
plt.figure()
plt.imshow(annotations.get_frame(frame_num=55).show())
# Play video with the Dataloop video player
annotations.video_player()
def main():
"""
Convert yolo annotation and images to Dataloop
Yolo annotations format:
For each image there is a text file with same name that has a list of box location and label index
"""
import dtlpy as dl
from PIL import Image
import os
# Get project and dataset
project = dl.projects.get(project_name='Fruits')
dataset = project.datasets.get(dataset_name='Rambutan')
images_and_annotations_path = '/home/fruits/data'
# read all images from local dataset
img_filepaths = list()
for path, subdirs, files in os.walk(images_and_annotations_path):
for filename in files:
striped, ext = os.path.splitext(filename)
if ext in ['.jpeg']:
img_filepaths.append(os.path.join(path, filename))
classes_filepath = '/home/fruits/classes.txt'
# get labels from file
with open(classes_filepath, 'r') as f:
labels = {
i_line: label.strip()
for i_line, label in enumerate(f.readlines())
}
for filepath in img_filepaths:
# get image height and width
img = Image.open(filepath)
width, height = img.size()
# upload item to platform
item = dataset.items.update(filepath=filepath)
# get YOLO annotations
_, ext = os.path.splitext(filepath)
annotation_filepath = filepath.replace(ext, '.txt')
with open(annotation_filepath, 'r') as f:
annotations = f.read().split('\n')
builder = item.annotations.builder()
# convert to Dataloop annotations
for annotation in annotations:
if not annotation:
continue
# convert annotation format
elements = annotation.split(" ")
label_id = elements[0]
xmin_add_xmax = float(elements[1]) * (2.0 * float(width))
ymin_add_ymax = float(elements[2]) * (2.0 * float(height))
w = float(elements[3]) * float(width)
h = float(elements[4]) * float(height)
left = (xmin_add_xmax - w) / 2
top = (ymin_add_ymax - h) / 2
right = left + w
bottom = top + h
# add to annotations
builder.add(annotation_definition=dl.Box(top=top,
left=left,
bottom=bottom,
right=right,
label=labels[label_id]))
# upload all annotations of an item
builder.upload()
def main():
"""
Detect and track (using model and some tracker) and upload annotation to platform
:return:
"""
import cv2
import dtlpy as dl
##########################
# Load model and tracker #
##########################
# load your model for detection
model = load_some_model()
# load any tracking algorithm to track detected elements
tracker = load_some_tracker()
##############
# load video #
##############
video_path = 'some/video/path'
vid = cv2.VideoCapture(video_path)
if not vid.isOpened():
raise IOError("Couldn't open webcam or video")
video_fps = vid.get(cv2.CAP_PROP_FPS)
video_size = (int(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),
int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)))
video_frames = int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
############
# Platform #
###########
# get the item from platform
item = dl.projects.get(project_name='MyProject') \
.datasets.get(dataset_name='MyDataset') \
.items.get(filepath='/path/to/video.mp4')
builder = item.annotations.builder()
#######
# Run #
#######
frame_num = 0
while True:
# get new frame from video
return_value, frame = vid.read()
if not return_value:
break
# get detection
detections = model.predict(frame)
# update tracker
tracked_elements = tracker.update(detections, frame)
# update annotations object
for element in tracked_elements:
# element.bb - format of the bounding box is 2 points in 1 array - [x_left, y_top, x_right, y_bottom])
# tracking id of each element is in element.id. to keep the ids of the detected elements
left, top, bottom, right = element.bb, # points bounding box annotation
builder.add(annotation_definition=dl.Box(top=top,
left=left,
right=right,
bottom=bottom,
label=element.label),
object_id=element.id,
frame_num=frame_num)
# increase frame number
frame_num += 1
if cv2.waitKey(1) & 0xFF == ord('q'):
break
##################################
# Upload annotations to platform #
##################################
item.annotations.upload(builder.to_platform())
def track_bounding_box(self,
item,
annotation,
frame_duration=100,
progress=None):
"""
:param item: dl.Item
:param annotation: dl.Annotation
:param frame_duration: How many frames ahead to track
:param progress:
:return:
"""
try:
assert isinstance(item, dl.Item)
assert isinstance(annotation, dl.Annotation)
item_stream_url = item.stream
bbx = annotation.coordinates
start_frame = annotation.start_frame
annotation.fps = item.fps
logger.info('[Tracker] Started')
tic_get_cap = time.time()
logger.info('[Tracker] video url: {}'.format(item_stream_url))
cap = cv2.VideoCapture('{}?jwt={}'.format(item_stream_url,
dl.token()))
runtime_get_cap = time.time() - tic_get_cap
if not cap.isOpened():
logger.error('[Tracker] failed opening video url')
raise ValueError('cant open video stream. item id: {}'.format(
item_stream_url))
logger.info('[Tracker] received bbs(xyxy): {}'.format(bbx))
tic_first_frame_set = time.time()
cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
runtime_first_frame_set = time.time() - tic_first_frame_set
mask_enable = False
runtime_load_frame = list()
runtime_track = list()
tic_total = time.time()
state_bbx = None
for i_frame in range(frame_duration):
logger.info(
'[Tracker] processing frame #{}'.format(start_frame +
i_frame))
tic = time.time()
ret, frame = cap.read()
if not ret:
break
runtime_load_frame.append(time.time() - tic)
tic = time.time()
# get bounding box
top = bbx[0]['y']
left = bbx[0]['x']
bottom = bbx[1]['y']
right = bbx[1]['x']
w = right - left
h = bottom - top
if i_frame == 0: # init
target_pos = np.array([left + w / 2, top + h / 2])
target_sz = np.array([w, h])
# init tracker
state_bbx = siamese_init(im=frame,
target_pos=target_pos,
target_sz=target_sz,
model=self.siammask,
hp=self.cfg['hp'],
device=self.device)
else:
# track
state_bbx = siamese_track(state=state_bbx,
im=frame,
mask_enable=mask_enable,
refine_enable=True,
device=self.device)
if mask_enable:
state_bbx['ploygon'].flatten()
mask = state_bbx['mask'] > state_bbx['p'].seg_thr
annotation.add_frame(annotation_definition=dl.Box(
top=np.min(np.where(mask)[0]),
left=np.min(np.where(mask)[1]),
bottom=np.max(np.where(mask)[0]),
right=np.max(np.where(mask)[1]),
label=annotation.label),
frame_num=start_frame + i_frame)
else:
top = state_bbx['target_pos'][
1] - state_bbx['target_sz'][1] / 2
left = state_bbx['target_pos'][
0] - state_bbx['target_sz'][0] / 2
bottom = state_bbx['target_pos'][
1] + state_bbx['target_sz'][1] / 2
right = state_bbx['target_pos'][
0] + state_bbx['target_sz'][0] / 2
annotation.add_frame(annotation_definition=dl.Box(
top=top,
left=left,
bottom=bottom,
right=right,
label=annotation.label),
frame_num=start_frame + i_frame)
runtime_track.append(time.time() - tic)
runtime_total = time.time() - tic_total
fps = frame_duration / runtime_total
logger.info('[Tracker] Finished.')
logger.info('[Tracker] Runtime information: \n'
'Total runtime: {:.2f}[s]\n'
'FPS: {:.2f}fps\n'
'Get url capture object: {:.2f}[s]\n'
'Initial set frame: {:.2f}[s]\n'
'Total track time: {:.2f}[s]\n'
'Mean load per frame: {:.2f}\n'
'Mean track per frame: {:.2f}'.format(
runtime_total, fps, runtime_get_cap,
runtime_first_frame_set,
np.sum(runtime_load_frame) + np.sum(runtime_track),
np.mean(runtime_load_frame),
np.mean(runtime_track)))
annotation.delete()
item.annotations.upload(annotations=[annotation])
progress.update(status='success')
except Exception:
logger.exception('Failed during track:')
raise