doi

DOI - Detection of Objects in Images

DOI is an application that searches for images in a provided folder and detects objects in the images.

This is the stand alone application, in Python 3.

The main inputs are a directory, to search for images, and the object classes that you want to detect (see Usage). The result is a json file with each class of objects selected and copies of the images with the objects detected surrounded by boxes.

This application was built in Python 3.7.

It only supports Windows (for now).

There are an ingest module and a report module to use this application in Autopsy. Check it out in this repository.

Installation

• Install Python version 3 (tested in Python 3.7)
• Install CUDA (optional, only with you have a compatible GPU)
• Run the command: pip install numpy opencv-python pillow
• Put weights files in folder config/weights. Download at least one from:
  • http://pjreddie.com/media/files/yolov3.weights (recomended, for default configuration)
  • http://pjreddie.com/media/files/yolov3-tiny.weights (for tiny configuration)
  • http://pjreddie.com/media/files/yolov3-openimages.weights (for openimages configuration)

Usage

Detector

Get the list of images with dogs and horses (-cls or --classes), from all the images in directory and sub-directories:

python doi.py detect <directory_with_images> --classes dog horse

Get the list of images with dogs and baseball bats, from all the images in directory and sub-directories ("baseball bat" has a space between, so you must use quotes):

python doi.py detect <directory_with_images> --classes dog "baseball bat"

Using other configuration (-c or --config):

python doi.py --config tiny detect <directory_with_images> --classes dog horse

Other options as CPU mode (-ng or --nogpu) and threshold of 0.25 (-t or --threshold):

python doi.py --config tiny detect <directory_with_images> --classes dog horse --nogpu --threshold 0.25

Using a json file as input for classes to search for (-clsf or --classesfile):

python doi.py detect <directory_with_images> --classesfile <file_with_the_classes_to_search_for.json>

Do not clean images in results directory, created in a previous detection (-nrc or --no-results-clean) (by default, it cleans the results directory before saving new files):

python doi.py detect <directory_with_images> --classes dog horse --no-results-clean

Disable saving images with the detections in results directory (-nout or --no-output-images):

python doi.py detect <directory_with_images> --classes dog horse --no-output-images

Get supported classes

Get supported object classes (it will write a json file in results folder with the supported classes, that can be used as template for a json file with the classes to search for):

python doi.py classes

Get supported object classes for configuration (-c or --config):

python doi.py --config tiny classes

Example of a json file for indicating the classes to search for (detector will search for persons and dogs):

{
    "person": true,
    "dog": true,
    "car": false
}

About the detectors modules

darknet

• yolo_cpp_dll.dll and yolo_cpp_dll_nogpu.dll compiled from https://github.com/AlexeyAB/darknet
• pthreadGC2.dll and pthreadVC2.dll from https://github.com/AlexeyAB/darknet/tree/master/build/darknet/x64
• __init__.py is the same as darknet.py from https://github.com/AlexeyAB/darknet/blob/master/darknet.py

opencv

• Using opencv-python (dnn module), based on the example from https://github.com/arunponnusamy/object-detection-opencv

To Build EXE

You can build the executable file of this application (or you can use the one that is with the Autopsy ingest module). With the EXE, there is no need to install Python or other dependencies, but the folder 'config' must be in same directory of the EXE file.

• Install pyinstaller: pip install pyinstaller
• In the main project directory, run the command: pyinstaller doi.spec --distpath .
• To execute the application in exe mode, the commands and options are the same, eg.: doi detect <directory_with_images> --classes dog horse

Issues

• This application does not support directories with non-ASCII characters, due to its dependencies, namely OpenCV and Darknet.