pyDarknet

This is Python wrapper library for Darknet which is a deep neural network framework by AlexeyAB running YOLO object detector.

This is implemented by ctypes and you can use complied shared library file (libdark.so / darknet.dll) on Darknet directely in you Python project.

Requirements

• opencv-python

pip install -r requirements.txt

Usages

How to import pyDarknet on your project.

1. Copy compiled shared library file to on your project.
2. Rename the shared library file libdarknet-gpu.so or libdarknet.so.
   • pyDarknet selects libdarknet-gpu.so supporting GPU acceleration first. If there's no library surpoting GPU acceleration, pyDarknet select libdarknet.so alternatively.
   • pyDarknet can detect shared library extensions by operating systems.
     e.g.) libdarknet.so is for Linux, libdarknet.dll is for Windows.
3. Just import pydarknet package on your project.

   import pydarknet

How to load YOLOv4 by pyDarknet.

It's simple. Just create a Detector.

detector = pydarknet.Detector('yolov4.cfg', 'yolov4.weights', 'coco.names')

cfg_filename and weights_filename assigned as each yolov4.cfg and yolov4.weights must be filled. On the other hand, names_filename assigned as coco.names can be keep empty by passing None.

How to detect objects by YOLOv4 created by pyDarknet.

1. Prepare or load image.

   • The input must have HWC shape.
   • The input must be 3 channel BGR (NOT RGB) image commonly used in OpenCV.

   If you are using OpenCV mainly, you can use cv2.imread / cv2.imdecode.

   import cv2
   img = cv2.imread('image.png', cv2.IMREAD_COLOR)

2. Call detect method in Detector.

   bboxes = detector.detect(img)

   detect method has obj_thres, hier_thresh, nms_thresh and obj_filter.
   Three threshold paramaters having a posfix as _thresh are object detection paramaters. You can refence original paper or repository.

   obj_filter is an option filtering bounding boxes you want by object names. It is must be list of object names or indices.

   detector.detect(..., obj_filter=['car', 'bus'])

   or

   detector.detect(..., obj_filter['1', '2'])

Results of Detector

The detector returns list of tuples like (name, prob, gebox).

for name, prob, geobox in bboxes:
    x, y, w, h = geobox

What is geobox? To distinguish a bounding box including both name and geometry informations on image and a box including only geometry informations on image. I called the latter geobox.

name is detected object name defined in .names file. if you not set .names file in the detector, name is object index (NOT A NUMBER, BUT A STRING PRESENTATION NUMBER) defined in the model.