This framework can:
- parse all kinds of object detection datasets (ImageNet, COCO, YOLO, PascalVOC, OpenImage, CVAT, LabelMe, etc.) and show statistics,
- convert them to other formats (ImageNet, COCO, YOLO, PascalVOC, OpenImage, CVAT, LabelMe, etc.),
- and evaluate predictions using standard object detection metrics such as
$AP_{[.5:.05:.95]}$ ,$AP_{50}$ ,$mAP$ ,$AR_{1}$ ,$AR_{10}$ ,$AR_{100}$ .
This framework can be used both as a library in your own code and as a command line tool. This tool is designed to be simple to use, fast and correct.
You can install the package using pip:
pip install globox
The library has three main components:
BoundingBox
: represents a bounding box with a label and an optional confidence scoreAnnotation
: represent the bounding boxes annotations for one imageAnnotationSet
: represents annotations for a set of images (a database)
The AnnotationSet
class contains static methods to read different dataset formats:
# COCO
coco = AnnotationSet.from_coco(file_path="path/to/file.json")
# YOLOv5
yolo = AnnotationSet.from_yolo_v5(
folder="path/to/files/",
image_folder="path/to/images/"
)
# Pascal VOC
pascal = AnnotationSet.from_pascal_voc(folder="path/to/files/")
Annotation
offers file-level granularity for compatible datasets:
annotation = Annotation.from_labelme(file_path="path/to/file.xml")
For more specific implementations the BoundingBox
class contains lots of utilities to parse bounding boxes in different formats, like the create()
method.
AnnotationsSets
are set-like objects. They can be combined and annotations can be added:
gts = coco | yolo
gts.add(annotation)
Iterators and efficient lookup by image_id
's are easy to use:
if annotation in gts:
print("This annotation is present.")
if "image_123.jpg" in gts.image_ids:
print("Annotation of image 'image_123.jpg' is present.")
for box in gts.all_boxes:
print(box.label, box.area, box.is_ground_truth)
for annotation in gts:
nb_boxes = len(annotation.boxes)
print(f"{annotation.image_id}: {nb_boxes} boxes")
Datasets stats can printed to the console:
coco_gts.show_stats()
Database Stats
┏━━━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━┓
┃ Label ┃ Images ┃ Boxes ┃
┡━━━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━┩
│ aeroplane │ 10 │ 15 │
│ bicycle │ 7 │ 14 │
│ bird │ 4 │ 6 │
│ boat │ 7 │ 11 │
│ bottle │ 9 │ 13 │
│ bus │ 5 │ 6 │
│ car │ 6 │ 14 │
│ cat │ 4 │ 5 │
│ chair │ 9 │ 15 │
│ cow │ 6 │ 14 │
│ diningtable │ 7 │ 7 │
│ dog │ 6 │ 8 │
│ horse │ 7 │ 7 │
│ motorbike │ 3 │ 5 │
│ person │ 41 │ 91 │
│ pottedplant │ 6 │ 7 │
│ sheep │ 4 │ 10 │
│ sofa │ 10 │ 10 │
│ train │ 5 │ 6 │
│ tvmonitor │ 8 │ 9 │
├─────────────┼────────┼───────┤
│ Total │ 100 │ 273 │
└─────────────┴────────┴───────┘
Datasets can be converted to and saved in other formats:
# ImageNet
gts.save_imagenet(save_dir="pascalVOC_db/")
# YOLO Darknet
gts.save_yolo_darknet(
save_dir="yolo_train/",
label_to_id={"cat": 0, "dog": 1, "racoon": 2}
)
# YOLOv5
gts.save_yolo_v5(
save_dir="yolo_train/",
label_to_id={"cat": 0, "dog": 1, "racoon": 2},
)
# CVAT
gts.save_cvat(path="train.xml")
COCO Evaluation is also supported:
evaluator = COCOEvaluator(
ground_truths=gts,
predictions=dets
)
ap = evaluator.ap()
ar_100 = evaluator.ar_100()
ap_75 = evaluator.ap_75()
ap_small = evaluator.ap_small()
...
All COCO standard metrics can be displayed in a pretty printed table with:
evaluator.show_summary()
which outputs:
COCO Evaluation
┏━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━┳...┳━━━━━━━━┳━━━━━━━━┳━━━━━━━━┓
┃ Label ┃ AP 50:95 ┃ AP 50 ┃ ┃ AR S ┃ AR M ┃ AR L ┃
┡━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━╇...╇━━━━━━━━╇━━━━━━━━╇━━━━━━━━┩
│ airplane │ 22.7% │ 25.2% │ │ nan% │ 90.0% │ 0.0% │
│ apple │ 46.4% │ 57.4% │ │ 48.5% │ nan% │ nan% │
│ backpack │ 54.8% │ 85.1% │ │ 100.0% │ 72.0% │ 0.0% │
│ banana │ 73.6% │ 96.4% │ │ nan% │ 100.0% │ 70.0% │
. . . . . . . .
. . . . . . . .
. . . . . . . .
├───────────┼──────────┼────────┼...┼────────┼────────┼────────┤
│ Total │ 50.3% │ 69.7% │ │ 65.4% │ 60.3% │ 55.3% │
└───────────┴──────────┴────────┴...┴────────┴────────┴────────┘
The array of results can be saved in CSV format:
evaluator.save_csv("where/to/save/results.csv")
Custom evaluations can be achieved with:
evaluation = evaluator.evaluate(
iou_threshold=0.33,
max_detections=1_000,
size_range=(0.0, 10_000)
)
ap = evaluation.ap()
cat_ar = evaluation["cat"].ar
Evaluations are cached by (iou_threshold, max_detections, size_range)
keys. This means that repetead queries to the evaluator are fast!
If you only need to use Globox from the command line like an application, you can install the package through pipx:
pipx install globox
Globox will then be in your shell path and usable from anywhere.
Get a summary of annotations for one dataset:
globox summary /yolo/folder/ --format yolo
Convert annotations from one format to another one:
globox convert input/yolo/folder/ output_coco_file_path.json --format yolo --save_fmt coco
Evaluate a set of detections with COCO metrics, display them and save them in a CSV file:
globox evaluate groundtruths/ predictions.json --format yolo --format_dets coco -s results.csv
Show the help message for an exhaustive list of options:
globox summary -h
globox convert -h
globox evaluate -h
Clone the repo with its test data:
git clone https://github.com/laclouis5/globox --recurse-submodules=tests/globox_test_data
cd globox
Install with Poetry:
poetry install
Run the tests:
poetry run pytest
Speed benchmark can be executed with:
python tests/benchmark.py -n 5
The following speed test is performed using Python 3.11 and timeit
with 5 iterations on a 2021 MacBook Pro 14" (M1 Pro 8 Cores and 16 GB of RAM). The dataset is COCO 2017 Validation which comprises 5k images and 36 781 bounding boxes.
Task | COCO | CVAT | OpenImage | LabelMe | PascalVOC | YOLO | TXT |
---|---|---|---|---|---|---|---|
Parsing | 0.22s | 0.12s | 0.44s | 0.60s | 0.97s | 1.45s | 1.12s |
Saving | 0.32s | 0.17s | 0.14s | 1.06s | 1.08s | 0.91s | 0.85s |
AnnotationSet.show_stats()
: 0.02 s- Evalaution: 0.30 s
- Basic data structures and utilities
- Parsers (ImageNet, COCO, YOLO, Pascal, OpenImage, CVAT, LabelMe)
- Parser tests
- Database summary and stats
- Database converters
- Visualization options
- COCO Evaluation
- Tests with a huge load (5k images)
- CLI interface
- Make
image_size
optional and raise err when required (bbox conversion) - Make file saving atomic with a temporary to avoid file corruption
- Pip package!
- PascalVOC Evaluation
- Parsers for TFRecord and TensorFlow
- UI interface?
This repo is based on the work of Rafael Padilla.
Feel free to contribute, any help you can offer with this project is most welcome.