An open-source tool to seamlessly convert between popular computer vision label formats.
Popular label formats are sparsely documented and store different information. Understanding them and dealing with the differences is tedious and time-consuming. Labelformat aims to solve this pain.
- object-detection
- instance-segmentation
- Support for common dataset label formats (more coming soon)
- Support for common tool formats (more coming soon)
- Minimal dependencies, targets python 3.7 or higher
- Memory concious - datasets are processed file-by-file instead of loading everything in memory (when possible)
- Typed
- Tested with round trip tests to ensure consistency
- MIT license
This package is compatible with the following platforms:
- Windows
- macOS
- Linux
Note Labelformat is a young project, contributions and bug reports are welcome. Please see Contributing section below.
pip install labelformatConvert instance segmentation labels from COCO to YOLOv8:
labelformat convert \
--task instance-segmentation \
--input-format coco \
--input-file coco-labels/train.json \
--output-format yolov8 \
--output-file yolo-labels/data.yaml \
--output-split trainConvert object detection labels from KITTI to PascalVOC:
labelformat convert \
--task object-detection \
--input-format kitti \
--input-folder kitti-labels/labels \
--category-names cat,dog,fish \
--images-rel-path ../images \
--output-format pascalvoc \
--output-folder pascalvoc-labelsConvert object detection labels from Labelbox to Lightly:
labelformat convert \
--task object-detection \
--input-format labelbox \
--input-file labelbox-labels/export-result.ndjson \
--category-names cat,dog,fish \
--output-format lightly \
--output-folder lightly-labels/annotation-taskList the available tasks with:
$ labelformat convert --help
usage: labelformat convert [-h] --task
{instance-segmentation,object-detection}
Convert labels from one format to another.
optional arguments:
-h, --help
--task {instance-segmentation,object-detection}List the available formats for a given task with:
$ labelformat convert --task object-detection --help
usage: labelformat convert [-h] --task
{instance-segmentation,object-detection}
--input-format
{coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
--output-format
{coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
Convert labels from one format to another.
optional arguments:
-h, --help
--task {instance-segmentation,object-detection}
--input-format {coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
Input format
--output-format {coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
Output formatSpecify the input and output format to get required options for specific formats:
$ labelformat convert \
--task object-detection \
--input-format coco \
--output-format yolov8 \
--help
usage: labelformat convert [-h] --task
{instance-segmentation,object-detection}
--input-format
{coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
--output-format
{coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
--input-file INPUT_FILE --output-file OUTPUT_FILE
[--output-split OUTPUT_SPLIT]
Convert labels from one format to another.
optional arguments:
-h, --help
--task {instance-segmentation,object-detection}
--input-format {coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
Input format
--output-format {coco,kitti,labelbox,lightly,pascalvoc,yolov5,yolov6,yolov7,yolov8}
Output format
'coco' input arguments:
--input-file INPUT_FILE
Path to input COCO JSON file
'yolov8' output arguments:
--output-file OUTPUT_FILE
Output data.yaml file
--output-split OUTPUT_SPLIT
Split to usePlease refer to the code for a full list of available classes.
from pathlib import Path
from labelformat.formats import COCOObjectDetectionInput, YOLOv8ObjectDetectionOutput
# Load the input labels
label_input = COCOObjectDetectionInput(
input_file=Path("coco-labels/train.json")
)
# Convert to output format and save
YOLOv8ObjectDetectionOutput(
output_file=Path("yolo-labels/data.yaml"),
output_split="train",
).save(label_input=label_input)We will walk through in detail how to convert object detection labels from COCO format to YOLOv8 format and the other way around.
Let's assume we have coco.json in the coco-labels directory with following contents:
{
"info": {
"description": "COCO 2017 Dataset",
"url": "http://cocodataset.org",
"version": "1.0",
"year": 2017,
"contributor": "COCO Consortium",
"date_created": "2017/09/01"
},
"licenses": [
{
"url": "http://creativecommons.org/licenses/by/2.0/",
"id": 4,
"name": "Attribution License"
}
],
"images": [
{
"file_name": "image1.jpg",
"height": 416,
"width": 640,
"id": 0,
"date_captured": "2013-11-18 02:53:27"
},
{
"file_name": "image2.jpg",
"height": 428,
"width": 640,
"id": 1,
"date_captured": "2016-01-23 13:56:27"
}
],
"annotations": [
{
"area": 421,
"iscrowd": 0,
"image_id": 0,
"bbox": [540, 295, 23, 18],
"category_id": 2,
"id": 1
},
{
"area": 695.1853359360001,
"iscrowd": 0,
"image_id": 0,
"bbox": [513, 271, 21, 33],
"category_id": 0,
"id": 2
},
{
"area": 27826,
"iscrowd": 0,
"image_id": 1,
"bbox": [268, 63, 94, 295],
"category_id": 2,
"id": 16
}
],
"categories": [
{
"supercategory": "animal",
"id": 0,
"name": "cat"
},
{
"supercategory": "animal",
"id": 1,
"name": "dog"
},
{
"supercategory": "animal",
"id": 2,
"name": "fish"
}
]
}Convert it to YOLOv8 format with the following command:
labelformat convert \
--task object-detection \
--input-format coco \
--input-file coco-labels/coco.json \
--output-format yolov8 \
--output-file yolo-from-coco-labels/data.yaml \
--output-split trainThis creates the following data structure with YOLOv8 labels:
yolo-from-coco-labels/
├── data.yaml
└── labels/
├── image1.txt
└── image2.txt
The contents of the created files will be:
# data.yaml
names:
0: cat
1: dog
2: fish
nc: 3
path: .
train: images
# image1.txt
2 0.86171875 0.7307692307692307 0.0359375 0.04326923076923077
0 0.81796875 0.6911057692307693 0.0328125 0.07932692307692307
# image2.txt
2 0.4921875 0.49182242990654207 0.146875 0.6892523364485982
Unlike COCO format, YOLO uses relative image coordinates. To convert from YOLO to COCO we therefore have to provide also input images. We prepare the following folder structure:
yolo-labels/
├── data.yaml
├── images/
| ├── image1.jpg
| └── image2.jpg
└── labels/
├── image1.txt
└── image2.txt
The file contents will be as above. The location of the image folder
is defined in data.yaml with the path (root path) and train field.
Note that YOLO format allows specifying different data folders for
train, val and test data splits, we chose to use train for our example.
To convert to COCO run the command below. Note that we specify --input-split train:
labelformat convert \
--task object-detection \
--input-format yolov8 \
--input-file yolo-labels/data.yaml \
--input-split train \
--output-format coco \
--output-file coco-from-yolo-labels/coco.jsonThe command will produce coco-from-yolo-labels/coco.json with the following contents:
{
"images": [
{
"id": 0,
"file_name": "image1.jpg",
"width": 640,
"height": 416
},
{
"id": 1,
"file_name": "image2.jpg",
"width": 640,
"height": 428
}
],
"categories": [
{
"id": 0,
"name": "cat"
},
{
"id": 1,
"name": "dog"
},
{
"id": 2,
"name": "fish"
}
],
"annotations": [
{
"image_id": 0,
"category_id": 2,
"bbox": [540.0, 295.0, 23.0, 18.0]
},
{
"image_id": 0,
"category_id": 0,
"bbox": [513.0, 271.0, 21.0, 33.0]
},
{
"image_id": 1,
"category_id": 2,
"bbox": [268.0, 63.0, 94.0, 295.0]
}
]
}Note that converting from COCO to YOLO and back loses some information since the intermediate format does not store all the fields.
If you encounter a bug or have a feature suggestion we will be happy if you file a GitHub issue.
We also welcome contributions, please submit a PR.
The library targets python 3.7 and higher. We use poetry to manage the development environment.
Here is an example development workflow:
# Create a virtual environment with development dependencies
poetry env use python3.7
poetry install
# Make changes
...
# Autoformat the code
poetry run make format
# Run tests
poetry run make all-checksLightly is a spin-off from ETH Zurich that helps companies build efficient active learning pipelines to select the most relevant data for their models.
You can find out more about the company and it's services by following the links below: