In [21]:
import os
import glob
import yaml
import cv2
import copy
import numpy as np
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt
from tqdm import tqdm
from PIL import Image
from random import shuffle
import ruamel.yaml
In [2]:
# Define data paths
DATA_FOLDER = os.path.join('.', 'data')
SIM_FOLDER = os.path.join(DATA_FOLDER, 'simulator')
ADDITIONAL_IMAGES_FOLDER = os.path.join(SIM_FOLDER, 'additional')
EXISTING_ANNOTATIONS_FILE = os.path.join(SIM_FOLDER, 'sim_data_annotations.yaml')
FASTER_RCNN_GRAPH = os.path.join('.', 'model', 'faster-rcnn_frozen.pb')
COCO_LABELS_FILE = os.path.join('.', 'labels', 'mscoco_label_map.pbtxt')
In [3]:
# Define constants
tl_labels = {"Green" : 1, "Red" : 2, "Yellow" : 3, "off" : 4}
tl_coco_class = 10
tl_detection_threshold = 0.80
image_h = 600
image_w = 800
In [13]:
# Read in all the image files.
sim_image_files = glob.glob(os.path.join(ADDITIONAL_IMAGES_FOLDER, '**','*.jpg'), recursive=True)
# Read in existing annotations
existing_annotations = yaml.load(open(EXISTING_ANNOTATIONS_FILE, 'r').read())
# Get classes for all images
sim_image_data = pd.DataFrame(sim_image_files, columns=['Path'])
sim_image_data['Class'] = sim_image_data.apply(lambda df: df['Path'].split(os.path.sep)[-2], axis=1)
# Boxes contain list of [ymin, xmin, ymax, xmax] boxes
sim_image_data['Boxes'] = None
In [5]:
# Load Faster-RCNN
detection_graph = tf.Graph()
with detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(FASTER_RCNN_GRAPH, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def, name='')
In [6]:
def load_image_into_numpy_array(image):
(im_width, im_height) = image.size
return np.array(image.getdata()).reshape((im_height, im_width, 3)).astype(np.uint8)
In [14]:
# Run inference on images
with detection_graph.as_default():
with tf.Session(graph=detection_graph) as sess:
# Definite input and output Tensors for detection_graph
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
# Each box represents a part of the image where a particular object was detected.
detection_boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
# Each score represent how level of confidence for each of the objects.
# Score is shown on the result image, together with the class label.
detection_scores = detection_graph.get_tensor_by_name('detection_scores:0')
detection_classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
for idx, row in tqdm(sim_image_data.iterrows(), total=len(sim_image_data)):
image = Image.open(row['Path'])
# the array based representation of the image will be used later in order to prepare the
# result image with boxes and labels on it.
image_np = load_image_into_numpy_array(image)
# Expand dimensions since the model expects images to have shape: [1, None, None, 3]
image_np_expanded = np.expand_dims(image_np, axis=0)
# Actual detection.
(boxes, scores, classes, num) = sess.run(
[detection_boxes, detection_scores, detection_classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
classes = np.squeeze(classes).astype(np.int32)
boxes = np.squeeze(boxes)
scores = np.squeeze(scores)
# Check if detections include traffic lights
traffic_lights = []
for obj_class, score, box in zip(classes, scores, boxes):
if obj_class == tl_coco_class:
if score is not None and score > tl_detection_threshold:
traffic_lights.append(box)
sim_image_data.at[idx, 'Boxes'] = traffic_lights
In [23]:
new_annotations = []
# Extract and normalize boxes in existing annotations
for sample in existing_annotations:
new_sample = {}
new_sample['filename'] = 'data/simulator/'+sample['filename']
new_sample['annotations'] = []
for annotation in sample['annotations']:
box = {}
box['xmin'] = annotation['xmin']/image_w
box['xmax'] = (annotation['xmin'] + annotation['x_width'])/image_w
box['ymin'] = annotation['ymin']/image_h
box['ymax'] = (annotation['ymin'] + annotation['y_height'])/image_h
box['class'] = annotation['class']
new_sample['annotations'].append(box)
new_annotations.append(new_sample)
# Grab all the annotations from the additional dataset
for idx, row in sim_image_data.iterrows():
new_sample = {}
new_sample['filename'] = row['Path'][2:].replace('\\', '/')
new_sample['annotations'] = []
for annotation in row['Boxes']:
box = {}
box['xmin'] = annotation[1]
box['xmax'] = annotation[3]
box['ymin'] = annotation[0]
box['ymax'] = annotation[2]
box['class'] = row['Class']
new_sample['annotations'].append(box)
new_annotations.append(new_sample)
In [30]:
existing_annotations[0]['annotations'][0]['x_width']
Out[30]:
In [24]:
# Write all the annotations to file
def float_representer(dumper, value):
text = '{0:.4f}'.format(value)
return dumper.represent_scalar(u'tag:yaml.org,2002:float', text)
yaml.add_representer(float, float_representer)
shuffle(new_annotations)
with open('data\\simulator\\final_annotations.yaml', 'w') as outfile:
yaml.dump(new_annotations, outfile)
In [239]:
def draw_boxes(img, boxes, color=(0, 0, 255), thick=3):
# Make a copy of the image
imcopy = np.copy(img)
image_height = imcopy.shape[0]
image_width = imcopy.shape[1]
# Iterate through the bounding boxes
for box in boxes:
# Draw a rectangle given bbox coordinates
cv2.rectangle(imcopy, (int(box['xmin']*image_width), int(box['ymin']*image_height)),
(int(box['xmax']*image_width), int(box['ymax']*image_height)), color, thick)
# Return the image copy with boxes drawn
return imcopy
In [297]:
def random_jitter_image(img, boxes, xlate_range):
img_np = load_image_into_numpy_array(img)
new_img = np.copy(img_np)
height, width, channels = img_np.shape
xlate_x = np.random.uniform(xlate_range) - xlate_range/2
xlate_y = np.random.uniform(xlate_range) - xlate_range/2
M_xlate = np.float32([[1,0,xlate_x],[0,1,xlate_y]])
#Transform image
img_np = cv2.warpAffine(img_np, M_xlate, (width, height))
#Transform boxes
for box in boxes:
box['xmin'] = box['xmin'] + xlate_x/width
box['ymin'] = box['ymin'] + xlate_y/height
box['xmax'] = box['xmax'] + xlate_x/width
box['ymax'] = box['ymax'] + xlate_y/height
return img_np, boxes
In [382]:
index = 50
test = Image.open(new_annotations[index]['filename'])
boxes = copy.deepcopy(new_annotations[index]['annotations'])
test_jitter, boxes_jitter = random_jitter_image(test, boxes, 20)
final_image = draw_boxes(test_jitter, boxes_jitter)
plt.imshow(final_image)
plt.show()
In [31]:
train_annotations_file = os.path.join(SIM_FOLDER, 'final_annotations.yaml')
loaded_annotations = yaml.load(open(train_annotations_file, 'rb').read())