Day 21 - Python embedded in AlexeyAB/darknet

Day 21 - Python embedded in AlexeyAB/darknet

Like the previous Joseph/darknet, AlexeyAB/darknet also provides a python interface, which can be called directly by Python developers. It is convenient to combine with the original Python code. The difference is that it is not specifically placed in the python directory, but directly in the root directory Day 20 - reexamine the accuracy of map, F1, IOU, precision recall This article introduces the file structure of AlexeyAB/darknet in detail. The following figure is the file structure diagram of YOLO with AlexeyAB/darknet version installed. Here, only the parts that need to be used are explained:

  • darknet: the main executable of AlexeyAB/darknet.
  •,, by calling, you can directly call some main functions of AlexeyAB/darknet in Python, such as training, testing, loading network structure, etc.
  • AlexeyAB/darknet uses the shared function library written in c language, which can be called by other programming languages.

Figure 1. File structure of YOLO in AlexeyAB/darknet version

Day 15 - description of YOLO related settings In this article, we will use the Joseph/darknet version of the identification model to build a self built dataset. Now copy AlexeyAV/darknet related files (darknet,, to this folder

Figure 2. Image recognition folder of self built dataset, introduced into AlexeyAB/darknet

Then create a file, import it into, and then specify the relevant configuration. The two files yolov3.cfg and are in Day 15 - description of YOLO related settings This article has a detailed description, and yolov3.backup is in Day 16 - perform image recognition training This article is generated by training.

import cv2
import darknet
import time

def image_detection(image_path, network, class_names, class_colors, thresh):
    width = darknet.network_width(network)
    height = darknet.network_height(network)
    darknet_image = darknet.make_image(width, height, 3)

    image = cv2.imread(image_path)
    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image_resized = cv2.resize(image_rgb, (width, height),

    darknet.copy_image_from_bytes(darknet_image, image_resized.tobytes())
    detections = darknet.detect_image(network, class_names, darknet_image, thresh=thresh)
    image = darknet.draw_boxes(detections, image_resized, class_colors)
    return cv2.cvtColor(image, cv2.COLOR_BGR2RGB), detections

# Specify the network structure configuration file yolov3.cfg, the self built dataset configuration file, and the pre trained weight file yolov3.backup
network, class_names, class_colors = darknet.load_network(

prev_time = time.time() # It is used to calculate the time to identify a picture
print('predicting...', prev_time)
# Carry out image recognition, and return the picture of the drawn box and the recognition results. The recognition results include labels, confidence and the coordinates of the box
image, detections = image_detection(
        './labels/00-frame-608x608-0030.jpg', network, class_names, class_colors, 0.25
# Print the label, confidence, and the coordinates of the box
darknet.print_detections(detections, '--ext_output')
# Display identification time
print((time.time() - prev_time))
# Write the result picture into the file result.jpg
cv2.imwrite('result.jpg', image)

Before running, you need to install OpenCV, because this program will be used. The screen is as follows. Three objects are identified, followed by the coordinates of the upper left corner and the length and width. The time is about 0.566 seconds.

pip3 install opencv-python

Figure 3. Text output screen executed by

Figure 4. Graphic output screen executed by

Compared with the original command mode, the following is the execution syntax of the command line.

./darknet detector test cfg/ cfg/yolov3.cfg weights/yolov3.backup ./labels/00-frame-608x608-0030.jpg -ext_output

Figure 5. Text output screen of AlexeyAB version instruction execution

Figure 6. Graphic output screen of AlexeyAB version instruction execution

It is found that the content of the object with edge is the same, but the confidence part, the part using python, is not rounded to three digits after the decimal point, so the results are 99.22%, 99.94%, 99.97%, while the instruction part is 99%, 100%, 100%; The identification time is also similar. The instruction execution time is about 0.558 seconds, while the python part is 0.566 seconds; The strangest thing is that the bbox data is different. The following table lists the comparison table of the coordinates of the box using Python and the box using Command. Directly look at the result picture. The box is the same. You can find that Python takes the coordinates of the center point and Command takes the coordinates of the upper left corner.

Table 1. Comparison of box coordinates using Python and Command


The following is the output coordinate content using Python and Command.

# using Pyhon
Altolamprologuscompressiceps: 99.22%    (left_x: 102   top_y:  420   width:   171   height:  93)
Altolamprologuscompressiceps: 99.94%    (left_x: 235   top_y:  373   width:   86   height:  51)
Altolamprologuscompressiceps: 99.97%    (left_x: 311   top_y:  400   width:   147   height:  77)

# using Command
Altolamprologuscompressiceps: 99%	(left_x:   16   top_y:  374   width:  171   height:   93)
Altolamprologuscompressiceps: 100%	(left_x:  192   top_y:  348   width:   86   height:   51)
Altolamprologuscompressiceps: 100%	(left_x:  238   top_y:  361   width:  147   height:   77)       

Use the following program to draw the box of the first object according to the coordinates provided by Command, as shown below.

import cv2

inputImgPath = '../labels/00-frame-608x608-0030.jpg'
outputImgPath = '../image/00-frame-608x608-0030-1.jpg'
img_w= img_h = 608
cv2image = cv2.imread(inputImgPath)

# Altolamprologuscompressiceps: 99%	(left_x:   16   top_y:  374   width:  171   height:   93)
min_x, min_y, max_x, max_y = 16,374,16+171,374+93
cv2.rectangle(cv2image, (int(min_x),int(min_y)), (int(max_x),int(max_y)), (0,255,255), 2)
cv2.imwrite(r'{}'.format(outputImgPath), cv2image)

Figure 7. Verifying the output coordinates of Python and Command

reference material

  • Yolo v4, v3 and v2 for Windows and Linux,

Tags: Python yolo

Posted on Fri, 29 Oct 2021 21:16:34 -0400 by scoppc