More efficient way of loading images for detection - tensorflow

Question

I am using tensorflow object detection api to do some semi real time object detection tasks.
The images will be taken by camera at a speed of 2 images/sec. Each image will be cropped into 4 small images so in total I need to process 8 images/sec.
My detection model has been exported into a frozen graph (.pb file) and loaded in GPU memory. Then I load images to numpy arrays to feed them into my model.
The detection itself only takes about 0.1 sec/image, however, loading each image takes about 0.45 sec.
The script I am using was revised from the code samples provided by object detection api(link), it reads each image and convert them into numpy array and then feed into detection models. The most time consumming part of this process is load_image_into_numpy_array, it takes almost 0.45 seconds.
The script is in below:
import numpy as np
import os
import six.moves.urllib as urllib
import sys
import tarfile
import tensorflow as tf
import zipfile
import timeit
import scipy.misc
from collections import defaultdict
from io import StringIO
from matplotlib import pyplot as plt
from PIL import Image
from utils import label_map_util
from utils import visualization_utils as vis_util
# Path to frozen detection graph. This is the actual model that is used for the
# object detection.
PATH_TO_CKPT = 'animal_detection.pb'
# List of the strings that is used to add correct label for each box.
PATH_TO_LABELS = os.path.join('data', 'animal_label_map.pbtxt')
NUM_CLASSES = 1
detection_graph = tf.Graph()
with detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def,name='')
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map,
max_num_classes=NUM_CLASSES,
use_display_name=True)
category_index = label_map_util.create_category_index(categories)
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)
# For the sake of simplicity we will use only 2 images:
# image1.jpg
# image2.jpg
# If you want to test the code with your images, just add path to the
# images to the TEST_IMAGE_PATHS.
PATH_TO_TEST_IMAGES_DIR = 'test'
TEST_IMAGE_PATHS = [
os.path.join(PATH_TO_TEST_IMAGES_DIR,'image{}.png'.format(i)) for i in range(1, 10) ]
# Size, in inches, of the output images.
IMAGE_SIZE = (12, 8)
config = tf.ConfigProto()
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
with detection_graph.as_default():
with tf.Session(graph=detection_graph, config=config) as sess:
for image_path in TEST_IMAGE_PATHS:
start = timeit.default_timer()
image = Image.open(image_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)
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
end = timeit.default_timer()
print(end-start)
start = timeit.default_timer()
# Each box represents a part of the image where a particular object was detected.
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.
scores = detection_graph.get_tensor_by_name('detection_scores:0')
classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
# Actual detection.
(boxes, scores, classes, num_detections) = sess.run(
[boxes, scores, classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
stop = timeit.default_timer()
print (stop - start)
# Visualization of the results of a detection.
vis_util.visualize_boxes_and_labels_on_image_array(
image_np,
np.squeeze(boxes),
np.squeeze(classes).astype(np.int32),
np.squeeze(scores),
category_index,
use_normalized_coordinates=True,
line_thickness=2)
I am thinking of a more efficient way to load images that are produced by camera, the first thought is to avoid numpy array and try to use tensorflow native ways to load images, but I have no idea where to get start since I am very new to tensorflow.
If I could find some tensorflow way to load images, maybe I could take 4 images into 1 batch and feed them into my model so that I might get some improvement in speed.
An immature idea is try to save 4 small images cropped from 1 raw image into a tf_record file, and load tf_record file as one batch to feed the model, but I have no idea how to achieve that.
Any help will be appreciated.

Answer 1

I found one solution that can reduce image loading from 0.4 second to 0.01 second. I will post answer here in case if someone also has same problem.
Instead of using PIL.Image and numpy, we could use imread in opencv.
I also managed to batch images so that we can achieve a better speedup.
The script goes as follow:
import numpy as np
import os
import six.moves.urllib as urllib
import sys
import tensorflow as tf
import timeit
import cv2
from collections import defaultdict
from utils import label_map_util
from utils import visualization_utils as vis_util
MODEL_PATH = sys.argv[1]
IMAGE_PATH = sys.argv[2]
BATCH_SIZE = int(sys.argv[3])
# Path to frozen detection graph. This is the actual model that is used for the
# object detection.
PATH_TO_CKPT = os.path.join(MODEL_PATH, 'frozen_inference_graph.pb')
# List of the strings that is used to add correct label for each box.
PATH_TO_LABELS = os.path.join('data', 'animal_label_map.pbtxt')
NUM_CLASSES = 1
detection_graph = tf.Graph()
with detection_graph.as_default():
od_graph_def = tf.GraphDef()
with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
serialized_graph = fid.read()
od_graph_def.ParseFromString(serialized_graph)
tf.import_graph_def(od_graph_def,name='')
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map,
max_num_classes=NUM_CLASSES,
use_display_name=True)
category_index = label_map_util.create_category_index(categories)
PATH_TO_TEST_IMAGES_DIR = IMAGE_PATH
TEST_IMAGE_PATHS = [
os.path.join(PATH_TO_TEST_IMAGES_DIR,'image{}.png'.format(i)) for i in range(1, 129) ]
config = tf.ConfigProto()
config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
with detection_graph.as_default():
with tf.Session(graph=detection_graph, config=config) as sess:
for i in range(0, len(TEST_IMAGE_PATHS), BATCH_SIZE):
images = []
start = timeit.default_timer()
for j in range(0, BATCH_SIZE):
image = cv2.imread(TEST_IMAGE_PATHS[i+j])
image = np.expand_dims(image, axis=0)
images.append(image)
image_np_expanded = np.concatenate(images, axis=0)
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.
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.
scores = detection_graph.get_tensor_by_name('detection_scores:0')
classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
# Actual detection.
(boxes, scores, classes, num_detections) = sess.run(
[boxes, scores, classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
stop = timeit.default_timer()
print (stop - start)