Computer Vision Utilities

Kano currently offers classes for handling YOLO-formatted datasets, one of the most commonly used formats in Object Detection tasks. With the YoloImage and YoloDataset classes, developers can load, explore, merge, split, and rename their datasets. Here are some examples:

Test demo with Google Colab here:

YOLO-formatted dataset tasks

Note

A valid YOLO-formatted dataset will have a folder tree like this

    dataset_name:
    ├─ train
    |  ├─ images
    |  |  ├─ image1.jpg
    |  |  └─ ...
    |  └─ labels
    |     ├─ image1.txt
    |     └─ ...
    ├─ valid (optional)
    ├─ test (optional)
    └─ data.yaml (must include a list of classes names inside "names" field)

Load and show a YOLO image

from kano.dataset_utils import YoloImage


label_dict = {0: "dog"}
image_path = "/content/dog_detection/train/images/dog_1_jpg.rf.dc6fff7fee9b6fb637b890b8387c6ce0.jpg"
image = YoloImage(image_path, label_dict)
image.show_image()
image.show_annotated_image()

# or you can get the annotated image and show
from kano.image import show_image


annotated_image = image.get_annotated_image()
show_image(annotated_image)

Result:

Original image:

Annotated image:

Load and show summary of a YOLO dataset

from kano.dataset_utils import YoloDataset


dataset_path = "/content/cat_detection"
dataset = YoloDataset(dataset_path)

# print summary information
dataset.summary()

# plot sample images
dataset.show_sample()

Result:

Summary dataset cat_detection:
- Classes:  ['cat']
- Subsets:
  + train: 1 images
  + test: 1 images
- Total images: 2

Merge datasets

When merging datasets, the names are also merged and reindexed alphabetically. The training images will be saved in the train folder, and similarly for the validation and test folders.

from kano.dataset_utils import YoloDataset


datasets_paths = ["cat_detection", "dog_detection"]
merged_dataset_path = "animals_detection"

YoloDataset.merge_datasets(datasets_paths, merged_dataset_path)

YoloDataset(merged_dataset_path).show_sample()

Result:

Input datasets:
Summary dataset cat_detection:
- Classes:  ['cat']
- Subsets:
  + train: 1 images
  + test: 1 images
- Total images: 2
Summary dataset dog_detection:
- Classes:  ['dog']
- Subsets:
  + train: 1 images
  + valid: 1 images
- Total images: 2
Merged dataset:
Summary dataset animals_detection:
- Classes:  ['cat', 'dog']
- Subsets:
  + train: 2 images
  + valid: 1 images
  + test: 1 images
- Total images: 4

Split train, validation, test datasets

from kano.dataset_utils import YoloDataset


dataset = YoloDataset("animals_detection")
dataset.split("splitted_dataset", ratios=[0.4, 0.3])

Result:

Summary dataset animals_detection:
- Classes:  ['cat', 'dog']
- Subsets:
  + train: 2 images
  + valid: 1 images
  + test: 1 images
- Total images: 4
Summary dataset splitted_dataset:
- Classes:  ['cat', 'dog']
- Subsets:
  + train: 1 images
  + valid: 1 images
  + test: 2 images
- Total images: 4

Rename classes

Classes can be removed by renaming to None.

from kano.dataset_utils import YoloDataset


renamed_dataset_path = "renamed_dataset"
renaming_dict = {"cat": "animal", "dog": None}

dataset = YoloDataset("animals_detection")
dataset.rename_classes(renamed_dataset_path, renaming_dict)

renamed_dataset = YoloDataset(renamed_dataset_path)
renamed_dataset.show_sample()

Result

Summary dataset animals_detection:
- Classes:  ['cat', 'dog']
- Subsets:
  + train: 2 images
  + valid: 1 images
  + test: 1 images
- Total images: 4
Classes after renaming: ['animal']
Summary dataset renamed_dataset:
- Classes:  ['animal']
- Subsets:
  + train: 2 images
  + valid: 1 images
  + test: 1 images
- Total images: 4

Object Detection tasks

Draw bounding box

You can add label for the bounding box. Current supported types of box are xyxy, xywh, s_xywh (scaled xywh).

import cv2
from kano.image import show_image
from kano.detect_utils import draw_bbox


image_path = "/content/dog_detection/train/images/dog_1_jpg.rf.dc6fff7fee9b6fb637b890b8387c6ce0.jpg"
image = cv2.imread(image_path)

label = "a cute dog"
bbox = [112, 326, 980, 827]

annotated_image = draw_bbox(image, bbox, bbox_type="xyxy", label=label, bbox_color=(0, 0, 255))
show_image(annotated_image)

Result: