OpenCV 'dnn' with NVIDIA GPUs: 1549% faster YOLO, SSD, and Mask R-CNN

In this tutorial, you’ll learn how to use OpenCV’s “dnn” module with an NVIDIA GPU for up to 1,549% faster object detection (YOLO and SSD) and instance segmentation (Mask R-CNN).

Last week, we discovered how to configure and install OpenCV and its “deep neural network” (dnn) module for inference using an NVIDIA GPU.

Using OpenCV’s GPU-optimized dnn module we were able to push a given network’s computation from the CPU to the GPU in only three lines of code:

# load the model from disk and set the backend target to a
# CUDA-enabled GPU
net = cv2.dnn.readNetFromCaffe(args["prototxt"], args["model"])
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

Today we’re going to discuss complete code examples in more detail — and by the end of the tutorial, you’ll be able to apply:

Single Shot Detectors (SSDs) at 65.90 FPS
YOLO object detection at 11.87 FPS
Mask R-CNN instance segmentation at 11.05 FPS

To learn how to use OpenCV’s dnn module and an NVIDIA GPU for faster object detection and instance segmentation, just keep reading!

Looking for the source code to this post?

OpenCV ‘dnn’ with NVIDIA GPUs: 1,549% faster YOLO, SSD, and Mask R-CNN

Inside this tutorial you’ll learn how to implement Single Shot Detectors, YOLO, and Mask R-CNN using OpenCV’s “deep neural network” (dnn) module and an NVIDIA/CUDA-enabled GPU.

Compile OpenCV’s ‘dnn’ module with NVIDIA GPU support

**Figure 1:** Compiling OpenCV’s DNN module with the CUDA backend allows us to perform object detection with YOLO, SSD, and Mask R-CNN deep learning models much faster.

If you haven’t yet, make sure you carefully read last week’s tutorial on configuring and installing OpenCV with NVIDIA GPU support for the “dnn” module — following that tutorial is an absolute prerequisite for this tutorial.

If you do not install OpenCV with NVIDIA GPU support enabled, OpenCV will still use your CPU for inference; however, if you try to pass the computation to the GPU, OpenCV will error out.

Project Structure

Before we review the structure of today’s project, grab the code and model files from the “Downloads” section of this blog post.

From there, unzip the files and use the tree command in your terminal to inspect the project hierarchy:

$ tree --dirsfirst
.
├── example_videos
│   ├── dog_park.mp4
│   ├── guitar.mp4
│   └── janie.mp4
├── opencv-ssd-cuda
│   ├── MobileNetSSD_deploy.caffemodel
│   ├── MobileNetSSD_deploy.prototxt
│   └── ssd_object_detection.py
├── opencv-yolo-cuda
│   ├── yolo-coco
│   │   ├── coco.names
│   │   ├── yolov3.cfg
│   │   └── yolov3.weights
│   └── yolo_object_detection.py
├── opencv-mask-rcnn-cuda
│   ├── mask-rcnn-coco
│   │   ├── colors.txt
│   │   ├── frozen_inference_graph.pb
│   │   ├── mask_rcnn_inception_v2_coco_2018_01_28.pbtxt
│   │   └── object_detection_classes_coco.txt
│   └── mask_rcnn_segmentation.py
└── output_videos

7 directories, 15 files

In today’s tutorial, we will review three Python scripts:

ssd_object_detection.py: Performs Caffe-based MobileNet SSD object detection on 20 COCO classes with CUDA.
yolo_object_detection.py: Performs YOLO V3 object detection on 80 COCO classes with CUDA.
mask_rcnn_segmentation.py: Performs TensorFlow-based Inception V2 segmentation on 90 COCO classes with CUDA.

Each of the model files and class name files are included in their respective folders with the exception of our MobileNet SSD (the class names are hardcoded in a Python list directly in the script). Let’s review the folder names in the order in which we’ll work with them today:

opencv-ssd-cuda/
opencv-yolo-cuda/
opencv-mask-rcnn-cuda/

As is evident by all three directory names, we will use OpenCV’s DNN module compiled with CUDA support. If your OpenCV is not compiled with CUDA support for your NVIDIA GPU, then you need to configure your system using the instructions in last week’s tutorial.

Implementing Single Shot Detectors (SSDs) using OpenCV’s NVIDIA GPU-Enabled ‘dnn’ module

**Figure 2:** Single Shot Detectors (SSDs) are known for being fast and efficient. In this tutorial, we’ll use Python + OpenCV + CUDA to perform even faster deep learning inference using an NVIDIA GPU.

The first object detector we’ll be looking at are Single Shot Detectors (SSDs), which we originally covered back in 2017:

Back then we could only run those SSDs on a CPU; however, today I’ll be showing you how to use your NVIDIA GPU to improve inference speed by up to 211%.

Open up the ssd_object_detection.py file in your project directory structure, and insert the following code:

# import the necessary packages
from imutils.video import FPS
import numpy as np
import argparse
import imutils
import cv2

# construct the argument parse and parse the arguments
ap = argparse.ArgumentParser()
ap.add_argument("-p", "--prototxt", required=True,
	help="path to Caffe 'deploy' prototxt file")
ap.add_argument("-m", "--model", required=True,
	help="path to Caffe pre-trained model")
ap.add_argument("-i", "--input", type=str, default="",
	help="path to (optional) input video file")
ap.add_argument("-o", "--output", type=str, default="",
	help="path to (optional) output video file")
ap.add_argument("-d", "--display", type=int, default=1,
	help="whether or not output frame should be displayed")
ap.add_argument("-c", "--confidence", type=float, default=0.2,
	help="minimum probability to filter weak detections")
ap.add_argument("-u", "--use-gpu", type=bool, default=False,
	help="boolean indicating if CUDA GPU should be used")
args = vars(ap.parse_args())

Here we’ve imported our packages. Notice that we do not require any special imports for CUDA. The CUDA capability is built in (via our compilation last week) to our cv2 import on Line 6.

Next let’s parse our command line arguments:

--prototxt: Our pretrained Caffe MobileNet SSD “deploy” prototxt file path.
--model: The path to our pretrained Caffe MobileNet SSD model.
--input: The optional path to our input video file. If it is not supplied, your first camera will be used by default.
--output: The optional path to our output video file.
--display: The optional boolean flag indicating whether we will diplay output frames to an OpenCV GUI window. Displaying frames costs CPU cycles, so for a true benchmark, you may wish to turn display off (by default it is on).
--confidence: The minimum probability threshold to filter weak detections. By default the value is set to 20%; however, you may override it if you wish.
--use-gpu: A boolean indicating whether the CUDA GPU should be used. By default this value is False (i.e., off). If you desire for your NVIDIA CUDA-capable GPU to be used for object detection with OpenCV, you need to pass a 1 value to this argument.

Next we’ll specify our classes and associated random colors:

# initialize the list of class labels MobileNet SSD was trained to
# detect, then generate a set of bounding box colors for each class
CLASSES = ["background", "aeroplane", "bicycle", "bird", "boat",
	"bottle", "bus", "car", "cat", "chair", "cow", "diningtable",
	"dog", "horse", "motorbike", "person", "pottedplant", "sheep",
	"sofa", "train", "tvmonitor"]
COLORS = np.random.uniform(0, 255, size=(len(CLASSES), 3))

And then we’ll load our Caffe-based model:

# load our serialized model from disk
net = cv2.dnn.readNetFromCaffe(args["prototxt"], args["model"])

# check if we are going to use GPU
if args["use_gpu"]:
	# set CUDA as the preferable backend and target
	print("[INFO] setting preferable backend and target to CUDA...")
	net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
	net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

As Line 35 indicates, we use OpenCV’s dnn module to load our Caffe object detection model.

A check is made to see if NVIDIA CUDA-enabled GPU should be used. From there, we set the backend and target accordingly (Lines 38-42).

Let’s go ahead and start processing frames and performing object detection with our GPU (provided the --use-gpu command line argument is turned on, of course):

# initialize the video stream and pointer to output video file, then
# start the FPS timer
print("[INFO] accessing video stream...")
vs = cv2.VideoCapture(args["input"] if args["input"] else 0)
writer = None
fps = FPS().start()

# loop over the frames from the video stream
while True:
	# read the next frame from the file
	(grabbed, frame) = vs.read()

	# if the frame was not grabbed, then we have reached the end
	# of the stream
	if not grabbed:
		break

	# resize the frame, grab the frame dimensions, and convert it to
	# a blob
	frame = imutils.resize(frame, width=400)
	(h, w) = frame.shape[:2]
	blob = cv2.dnn.blobFromImage(frame, 0.007843, (300, 300), 127.5)

	# pass the blob through the network and obtain the detections and
	# predictions
	net.setInput(blob)
	detections = net.forward()

	# loop over the detections
	for i in np.arange(0, detections.shape[2]):
		# extract the confidence (i.e., probability) associated with
		# the prediction
		confidence = detections[0, 0, i, 2]

		# filter out weak detections by ensuring the `confidence` is
		# greater than the minimum confidence
		if confidence > args["confidence"]:
			# extract the index of the class label from the
			# `detections`, then compute the (x, y)-coordinates of
			# the bounding box for the object
			idx = int(detections[0, 0, i, 1])
			box = detections[0, 0, i, 3:7] * np.array([w, h, w, h])
			(startX, startY, endX, endY) = box.astype("int")

			# draw the prediction on the frame
			label = "{}: {:.2f}%".format(CLASSES[idx],
				confidence * 100)
			cv2.rectangle(frame, (startX, startY), (endX, endY),
				COLORS[idx], 2)
			y = startY - 15 if startY - 15 > 15 else startY + 15
			cv2.putText(frame, label, (startX, y),
				cv2.FONT_HERSHEY_SIMPLEX, 0.5, COLORS[idx], 2)

Here we access our video stream. Note that the code is meant to be compatible with both video files and live video streams, which is why I elected not to use my threaded VideoStream class.

Looping over frames, we:

Read and preprocess incoming frames.
Construct a blob from the frame.
Detect objects using the Single Shot Detector and our GPU (if the --use-gpu flag was set).
Filter objects allowing only high --confidence objects to pass.
Annotate bounding boxes, class labels, and probabilities. If you need a refresher on OpenCV drawing basics, be sure to refer to my OpenCV Tutorial: A Guide to Learn OpenCV.

Finally, we’ll wrap up:

	# check to see if the output frame should be displayed to our
	# screen
	if args["display"] > 0:
		# show the output frame
		cv2.imshow("Frame", frame)
		key = cv2.waitKey(1) & 0xFF

		# if the `q` key was pressed, break from the loop
		if key == ord("q"):
			break

	# if an output video file path has been supplied and the video
	# writer has not been initialized, do so now
	if args["output"] != "" and writer is None:
		# initialize our video writer
		fourcc = cv2.VideoWriter_fourcc(*"MJPG")
		writer = cv2.VideoWriter(args["output"], fourcc, 30,
			(frame.shape[1], frame.shape[0]), True)

	# if the video writer is not None, write the frame to the output
	# video file
	if writer is not None:
		writer.write(frame)

	# update the FPS counter
	fps.update()

# stop the timer and display FPS information
fps.stop()
print("[INFO] elasped time: {:.2f}".format(fps.elapsed()))
print("[INFO] approx. FPS: {:.2f}".format(fps.fps()))

In the remaining lines, we:

Display the annotated video frames if required.
Capture key presses if we are displaying.
Write annotated output frames to a video file on disk.
Update, calculate, and print out FPS statistics.

Great job developing your SSD + OpenCV + CUDA script. In the next sections, we’ll analyze results using both our GPU and CPU.

Single Shot Detectors: 211% faster object detection with OpenCV’s ‘dnn’ module and an NVIDIA GPU

To see our Single Shot Detector in action, make sure you use the “Downloads” section of this tutorial to download (1) the source code and (2) pretrained models compatible with OpenCV’s dnn module.

From there, execute the following command to obtain a baseline for our SSD by running it on our CPU:

$ python ssd_object_detection.py \
	--prototxt MobileNetSSD_deploy.prototxt \
	--model MobileNetSSD_deploy.caffemodel \
	--input ../example_videos/guitar.mp4 \
	--output ../output_videos/ssd_guitar.avi \
	--display 0
[INFO] accessing video stream...
[INFO] elasped time: 11.69
[INFO] approx. FPS: 21.13

Here we are obtaining ~21 FPS on our CPU, which is quite good for an object detector!

To see the detector really fly, let’s supply the --use-gpu 1 command line argument, instructing OpenCV to push the dnn computation to our NVIDIA Tesla V100 GPU:

$ python ssd_object_detection.py \
	--prototxt MobileNetSSD_deploy.prototxt \
	--model MobileNetSSD_deploy.caffemodel \
	--input ../example_videos/guitar.mp4 \
	--output ../output_videos/ssd_guitar.avi \
	--display 0 \
	--use-gpu 1
[INFO] setting preferable backend and target to CUDA...
[INFO] accessing video stream...
[INFO] elasped time: 3.75
[INFO] approx. FPS: 65.90

Using our NVIDIA GPU, we’re now reaching ~66 FPS which improves our frames-per-second throughput rate by over 211%! And as the video demonstration shows, our SSD is quite accurate.

Note: As discussed by this comment by Yashas, the MobileNet SSD could perform poorly because cuDNN does not have optimized kernels for depthwise convolutions on all NVIDA GPUs. If you see your GPU results similar to your CPU results, this is likely the problem.

Implementing YOLO object detection for OpenCV’s NVIDIA GPU/CUDA-enabled ‘dnn’ module

**Figure 3:** YOLO is touted as being one of the fastest object detection architectures. In this section, we’ll use Python + OpenCV + CUDA to perform even faster YOLO deep learning inference using an NVIDIA GPU.

While YOLO is certainly one of the fastest deep learning-based object detectors, the YOLO model included with OpenCV is anything but — on a CPU, YOLO struggled to break 3 FPS.

Therefore, if you intend on using YOLO with OpenCV’s dnn module, you better be using a GPU.

Let’s take a look at how to use the YOLO object detector (yolo_object_detection.py) with OpenCV’s CUDA-enabled dnn module:

# import the necessary packages
from imutils.video import FPS
import numpy as np
import argparse
import cv2
import os

# construct the argument parse and parse the arguments
ap = argparse.ArgumentParser()
ap.add_argument("-y", "--yolo", required=True,
	help="base path to YOLO directory")
ap.add_argument("-i", "--input", type=str, default="",
	help="path to (optional) input video file")
ap.add_argument("-o", "--output", type=str, default="",
	help="path to (optional) output video file")
ap.add_argument("-d", "--display", type=int, default=1,
	help="whether or not output frame should be displayed")
ap.add_argument("-c", "--confidence", type=float, default=0.5,
	help="minimum probability to filter weak detections")
ap.add_argument("-t", "--threshold", type=float, default=0.3,
	help="threshold when applyong non-maxima suppression")
ap.add_argument("-u", "--use-gpu", type=bool, default=0,
	help="boolean indicating if CUDA GPU should be used")
args = vars(ap.parse_args())

Our imports are nearly the same as our previous script with one swap. In this script we don’t need imutils, but we do need Python’s os module for file I/O. Again, the CUDA capability is baked into our custom-compiled OpenCV installation.

Let’s review our command line arguments:

--yolo: The base path to your pretrained YOLO model directory.
--input: The optional path to our input video file. If it is not supplied, your first camera will be used by default.
--output: The optional path to our output video file.
--display: The optional boolean flag indicating whether we will use output frames to an OpenCV GUI window. Displaying frames costs CPU cycles, so for a true benchmark, you may wish to turn display off (by default it is on).
--confidence: The minimum probability threshold to filter weak detections. By default the value is set to 50%; however you may override it if you wish.
--threshold: The Non-Maxima Suppression (NMS) threshold is set to 30% by default.
--use-gpu: A boolean indicating whether the CUDA GPU should be used. By default this value is False (i.e., off). If you desire for your NVIDIA CUDA-capable GPU to be used for object detection with OpenCV, you need to pass a 1 value to this argument.

Next we’ll load our class labels and assign random colors:

# load the COCO class labels our YOLO model was trained on
labelsPath = os.path.sep.join([args["yolo"], "coco.names"])
LABELS = open(labelsPath).read().strip().split("\n")

# initialize a list of colors to represent each possible class label
np.random.seed(42)
COLORS = np.random.randint(0, 255, size=(len(LABELS), 3),
	dtype="uint8")

# derive the paths to the YOLO weights and model configuration
weightsPath = os.path.sep.join([args["yolo"], "yolov3.weights"])
configPath = os.path.sep.join([args["yolo"], "yolov3.cfg"])

# load our YOLO object detector trained on COCO dataset (80 classes)
print("[INFO] loading YOLO from disk...")
net = cv2.dnn.readNetFromDarknet(configPath, weightsPath)

# check if we are going to use GPU
if args["use_gpu"]:
	# set CUDA as the preferable backend and target
	print("[INFO] setting preferable backend and target to CUDA...")
	net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
	net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)

# determine only the *output* layer names that we need from YOLO
ln = net.getLayerNames()
ln = [ln[i[0] - 1] for i in net.getUnconnectedOutLayers()]

# initialize the width and height of the frames in the video file
W = None
H = None

# initialize the video stream and pointer to output video file, then
# start the FPS timer
print("[INFO] accessing video stream...")
vs = cv2.VideoCapture(args["input"] if args["input"] else 0)
writer = None
fps = FPS().start()

# loop over frames from the video file stream
while True:
	# read the next frame from the file
	(grabbed, frame) = vs.read()

	# if the frame was not grabbed, then we have reached the end
	# of the stream
	if not grabbed:
		break

	# if the frame dimensions are empty, grab them
	if W is None or H is None:
		(H, W) = frame.shape[:2]

	# construct a blob from the input frame and then perform a forward
	# pass of the YOLO object detector, giving us our bounding boxes
	# and associated probabilities
	blob = cv2.dnn.blobFromImage(frame, 1 / 255.0, (416, 416),
		swapRB=True, crop=False)
	net.setInput(blob)
	layerOutputs = net.forward(ln)

	# initialize our lists of detected bounding boxes, confidences,
	# and class IDs, respectively
	boxes = []
	confidences = []
	classIDs = []

	# loop over each of the layer outputs
	for output in layerOutputs:
		# loop over each of the detections
		for detection in output:
			# extract the class ID and confidence (i.e., probability)
			# of the current object detection
			scores = detection[5:]
			classID = np.argmax(scores)
			confidence = scores[classID]

			# filter out weak predictions by ensuring the detected
			# probability is greater than the minimum probability
			if confidence > args["confidence"]:
				# scale the bounding box coordinates back relative to
				# the size of the image, keeping in mind that YOLO
				# actually returns the center (x, y)-coordinates of
				# the bounding box followed by the boxes' width and
				# height
				box = detection[0:4] * np.array([W, H, W, H])
				(centerX, centerY, width, height) = box.astype("int")

				# use the center (x, y)-coordinates to derive the top
				# and and left corner of the bounding box
				x = int(centerX - (width / 2))
				y = int(centerY - (height / 2))

				# update our list of bounding box coordinates,
				# confidences, and class IDs
				boxes.append([x, y, int(width), int(height)])
				confidences.append(float(confidence))
				classIDs.append(classID)

	# apply non-maxima suppression to suppress weak, overlapping
	# bounding boxes
	idxs = cv2.dnn.NMSBoxes(boxes, confidences, args["confidence"],
		args["threshold"])

	# ensure at least one detection exists
	if len(idxs) > 0:
		# loop over the indexes we are keeping
		for i in idxs.flatten():
			# extract the bounding box coordinates
			(x, y) = (boxes[i][0], boxes[i][1])
			(w, h) = (boxes[i][2], boxes[i][3])

			# draw a bounding box rectangle and label on the frame
			color = [int(c) for c in COLORS[classIDs[i]]]
			cv2.rectangle(frame, (x, y), (x + w, y + h), color, 2)
			text = "{}: {:.4f}".format(LABELS[classIDs[i]],
				confidences[i])
			cv2.putText(frame, text, (x, y - 5),
				cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 2)

	# check to see if the output frame should be displayed to our
	# screen
	if args["display"] > 0:
		# show the output frame
		cv2.imshow("Frame", frame)
		key = cv2.waitKey(1) & 0xFF

		# if the `q` key was pressed, break from the loop
		if key == ord("q"):
			break

	# if an output video file path has been supplied and the video
	# writer has not been initialized, do so now
	if args["output"] != "" and writer is None:
		# initialize our video writer
		fourcc = cv2.VideoWriter_fourcc(*"MJPG")
		writer = cv2.VideoWriter(args["output"], fourcc, 30,
			(frame.shape[1], frame.shape[0]), True)

	# if the video writer is not None, write the frame to the output
	# video file
	if writer is not None:
		writer.write(frame)

	# update the FPS counter
	fps.update()

# stop the timer and display FPS information
fps.stop()
print("[INFO] elasped time: {:.2f}".format(fps.elapsed()))
print("[INFO] approx. FPS: {:.2f}".format(fps.fps()))

$ python yolo_object_detection.py --yolo yolo-coco \
	--input ../example_videos/janie.mp4 \
	--output ../output_videos/yolo_janie.avi \
	--display 0
[INFO] loading YOLO from disk...
[INFO] accessing video stream...
[INFO] elasped time: 51.11
[INFO] approx. FPS: 2.47

$ python yolo_object_detection.py --yolo yolo-coco \
	--input ../example_videos/janie.mp4 \
	--output ../output_videos/yolo_janie.avi \
	--display 0 \
	--use-gpu 1
[INFO] loading YOLO from disk...
[INFO] setting preferable backend and target to CUDA...
[INFO] accessing video stream...
[INFO] elasped time: 10.61
[INFO] approx. FPS: 11.87

Looking for the source code to this post?

OpenCV ‘dnn’ with NVIDIA GPUs: 1,549% faster YOLO, SSD, and Mask R-CNN

Compile OpenCV’s ‘dnn’ module with NVIDIA GPU support

Project Structure

Implementing Single Shot Detectors (SSDs) using OpenCV’s NVIDIA GPU-Enabled ‘dnn’ module

Single Shot Detectors: 211% faster object detection with OpenCV’s ‘dnn’ module and an NVIDIA GPU

Implementing YOLO object detection for OpenCV’s NVIDIA GPU/CUDA-enabled ‘dnn’ module

YOLO: 380% faster object detection with OpenCV’s NVIDIA GPU-enabled ‘dnn’ module

Implementing Mask R-CNN Instance Segmentation for OpenCV’s CUDA-Enabled ‘dnn’ module

Mask R-CNN: 1,549% faster Instance Segmentation with OpenCV’s ‘dnn’ NVIDIA GPU module

Making nearly any model compatible with OpenCV’s ‘dnn’ module run on an NVIDIA GPU

What's next? We recommend PyImageSearch University.

Summary

Download the Source Code and FREE 17-page Resource Guide

About the Author

PyImageSearch University

Finding extreme points in contours with OpenCV

Deep Learning on Amazon EC2 GPU with Python and nolearn

Understanding Tasks in Diffusers: Part 1

Topics

Books & Courses

PyImageSearch

Looking for the source code to this post?

OpenCV ‘dnn’ with NVIDIA GPUs: 1,549% faster YOLO, SSD, and Mask R-CNN

Compile OpenCV’s ‘dnn’ module with NVIDIA GPU support

Project Structure

Implementing Single Shot Detectors (SSDs) using OpenCV’s NVIDIA GPU-Enabled ‘dnn’ module

Single Shot Detectors: 211% faster object detection with OpenCV’s ‘dnn’ module and an NVIDIA GPU

Implementing YOLO object detection for OpenCV’s NVIDIA GPU/CUDA-enabled ‘dnn’ module

YOLO: 380% faster object detection with OpenCV’s NVIDIA GPU-enabled ‘dnn’ module

Implementing Mask R-CNN Instance Segmentation for OpenCV’s CUDA-Enabled ‘dnn’ module

Mask R-CNN: 1,549% faster Instance Segmentation with OpenCV’s ‘dnn’ NVIDIA GPU module

Making nearly any model compatible with OpenCV’s ‘dnn’ module run on an NVIDIA GPU

What's next? We recommend PyImageSearch University.

Summary

Download the Source Code and FREE 17-page Resource Guide

About the Author

Similar articles

You can learn Computer Vision, Deep Learning, and OpenCV.

Footer

Topics

Books & Courses

PyImageSearch

Access the code to this tutorial and all other 500+ tutorials on PyImageSearch

What's included in PyImageSearch University?