CVAT2YOLO Converter

Converter cvat dataset YOLO1.1 to YOLOv5 format

Installation

git clone https://github.com/ankhafizov/CVAT2YOLO.git

cd CVAT2YOLO

pip install -e .

Usage

1. If val and train parts splitted manually, folders train_folder and val_folder has to exist in cvat dataset folder:

cvat2yolo --cvat my_datset --mode manual --train_folder obj_Train_data --val_folder obj_Validation_data  --test_folder obj_Test_data --img_format jpg --output_folder my_dataset_yolov5

or more simplified (by default val_folder=obj_Validation_data, train_folder=my_obj_Train_data, test_folder=obj_Test_data, img_format=png will be taken):

cvat2yolo --cvat my_datset --mode manual --output_folder my_dataset_yolov5

if any of train, val or test folder does not exist, there will not be created corresponding folders in the output directory.

=====

2. In automatic mode ("autosplit") val and train folders would be merged and randomly splitted with the certain proportion:

cvat2yolo --cvat my_datset \
          --mode autosplit \
          --split 0.9 \
          --train_folder obj_Train_data \
          --val_folder obj_Validation_data \ 
          --test_folder obj_Test_data \ 
          --img_format png \
          --output_folder out/my_dataset_yolov5

or simlified:

cvat2yolo --cvat my_datset --mode autosplit --split 0.9 --output_folder out/my_dataset_yolov5

Help:

• --cvat - Path to the root folder of imported from cvat YOLO 1.1 dataset

• --mode - "autosplit" or "manual" (as it was exported from cvat, check the text above)

• --output_folder - Path to converted dataset folder {root}{dataset name}

• --split - A percentage of a split, e.g. 0.9 means split 0.9 for train and 0.1 for test (default None)

• --train_folder - Folder with Train subset inside cvat path (default obj_Train_data)

• --val_folder - Folder with Val subset inside cvat path (default obj_Validation_data)

• --test_folder - Folder with Test subset inside cvat path (default obj_Test_data)

• --img_format - File format of images (default png)

• --label_tfrms - Label union with another one existed in dataset (example: 'head->hood,helmet->hat' will unite 'head' with 'hood' and 'helmet' with 'hat' resulting 'hood' and 'hat' in output dataset)

• --percentage_empty - Percentage of images without any labels in relation to full dataset size (default =10, optional)

• --classes - Classes which labels to keep. So if in initiall dataset there are 3 classes (e.g. [A, B, C]), and there is flag --classes "A|C", only labels with classes A and C will be kept in output YOLOv5 dataset.

help:

cvat2yolo --help