Helper for dealing with MS-COCO annotations.
The MS COCO annotation format along with the pycocotools library is quite popular among the computer vision community. Yet I for one found it difficult to play around with the annotations. Deleting a specific category, combining multiple mini datasets to generate a larger dataset, viewing distribution of classes in the annotation file are things I would like to do without writing a separate script for each. The COCO Assistant is designed (or being designed) to assist with this problem. Please note that currently, the Assistant can only help out with object detection datasets. Any contributions and/or suggestions are welcome.
Your data directory should look as follows:
Example:
.
├── images
│ ├── train
│ ├── val
| ├── test
|
├── annotations
│ ├── train.json
│ ├── val.json
│ ├── test.json
pip install coco-assistant
# Clone the repository
git clone https://github.com/ashnair1/COCO-Assistant.git
# Build and install the library
make
Usage is similar to how you would use pycocotools
from coco_assistant import COCO_Assistant
# Specify image and annotation directories
img_dir = os.path.join(os.getcwd(), 'images')
ann_dir = os.path.join(os.getcwd(), 'annotations')
# Create COCO_Assistant object
cas = COCO_Assistant(img_dir, ann_dir)
The combine
function allows you to merge multiple datasets.
cas = COCO_Assistant(img_dir, ann_dir)
loading annotations into memory...
Done (t=0.09s)
creating index...
index created!
loading annotations into memory...
Done (t=0.06s)
creating index...
index created!
cas.combine()
Merging image dirs
100%|█████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 18.33it/s]
Merging annotations
100%|█████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 14.72it/s]
The merged dataset (images and annotation) can be found in ./results/combination
Removes a specific category from an annotation file.
cas = COCO_Assistant(img_dir, ann_dir)
loading annotations into memory...
Done (t=0.09s)
creating index...
index created!
loading annotations into memory...
Done (t=0.06s)
creating index...
index created!
cas.remove_cat()
['tiny.json', 'tiny2.json']
Select categories to use
tiny.json
Categories present:
['building', 'vehicles]
Select categories to use:
building
['building']
Press n if you're done entering categories, else continue
n
Using specified categories...
The modified annotation can be found in ./results/removal
- Generate countplot of instances per category that occur in the annotation files.
cas.ann_stats(stat="area",arearng=[10,144,512,1e5],save=False)
- Generate pie-chart that shows distribution of objects according to their size (as specified in areaRng).
cas.ann_stats(stat="cat", show_count=False, save=False)
Couldn't pycocotools
visualise annotations (via showAnns) as well? Sure it could, but I required a way to freely view all the annotations of a particular dataset so here we are.
cas.visualise()
Choose directory:
['tiny', 'tiny2']
- Converter for converting COCO annotations to YOLO format
- Write tests for stats, converters and visualiser.