Skip to content

Implemet the customization of yolov3 objectdetection algorithm based on darknet. Added the attribute classification based on the Pre_box. In addition, a variety of testing tools have been added.

Notifications You must be signed in to change notification settings

FlyingAnt2018/darknet-modify

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

darknet-modify

Introduction

This is a implemention of the customization yolov3 object detection algorithm based on darknet framework. Features are listed bollow.

  • 1st: Adding the Attributes classification based on the Pre_box, such as overlapping, head or tail, day or night, etc.
  • 2nd: Several image augmentation schemes have been added, such as: image block overlap, to make yolo higher recall. Ground glass effect is applied to make the edge of object more stable and steady.
  • 3rd: Based on the existing model, write the bounding boxes of the objects in a image into the xml file to assist further annotation. we provide all related source code, and corresponding executable file can be generated. For complete darknet code click here:https://pjreddie.com/darknet/install/
original paper:

Usage

  • To use Attribute function, please open macro "OPEN_OCC_CLASS_FLAG", in the project Scope.
  • To use model select function in the Test stage, set field "-test_mode" to 0, 1 or 2, mean run image in a folder to get pre_box, select models base their recall and precision, respectively.
  • To get objects' bounding box and save them to a xml file, please add field "-save_xml" in the script fild. eg:

@echo off
.\darknet_old.exe detector test E:/ E:\ E:\ -save_xml -test_mode 2 -dont_show
pause

About

Implemet the customization of yolov3 objectdetection algorithm based on darknet. Added the attribute classification based on the Pre_box. In addition, a variety of testing tools have been added.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published