jorgeuliana1 / train-detectron2

This repository is intended to contain scripts for training and testing neural networks using Detectron2.

SCRIPTS FOR TRAINING AND TESTING USING DETECTRON2

PURPOSE:

This repository is intended to contain scripts for training and testing neural networks using Detectron2 (facebookresearch/detectron2).

I created it to use in my paper, but feel free to use it too.

USAGE:

For training: python3 train.py dataset_info.json train_info.json

For testing: python3 test.py dataset_info.json test_info.json

Where dataset_info.json may look like:

{
    "TRAIN" : {
        "base_directory" : "foo/bar",
        "annotations_csv" : "foo/annotations.csv",
        "classes_json" : "foo/classes.json",
        "width" : 1280,
        "height" : 720
    },
    "TEST" : {
        "base_directory" : "foobar/bar",
        "annotations_csv" : "foobar/annotations.csv",
        "classes_json" : "foobar/classes.json",
        "width" : 1280,
        "height" : 720
    }
}

"TEST" and "TRAIN" tags refer to training and testing datasets respectively. annotations.csv file is something like:

image_001.jpg,0,0,200,200,CATEGORY_01
image_001.jpg,200,200,500,500,CATEGORY_02

in a general way it is:

image_path(relative to base_directory),x0,y0,x1,y1,category

x0, y0, x1, y1 follows the BoxMode.XYXY_ABS pattern.

classes.json file is something like:

{
    "__background__" : 0,
    "CATEGORY_01" : 1,
    "CATEGORY_02" : 2,
    "CATEGORY_03" : 3
}

where the "__background__" tag is optional.

The repository provides models for train_info.json and dataset_info.json.