https://github.com/davide710/augmentation
Twice the dataset, same the time for annotation!
https://github.com/davide710/augmentation
computer-vision dataannotations datalabeling linear-algebra machine-learning math object-detection opencv python yolo
Last synced: 28 days ago
JSON representation
Twice the dataset, same the time for annotation!
- Host: GitHub
- URL: https://github.com/davide710/augmentation
- Owner: davide710
- Created: 2024-02-01T16:57:43.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2025-01-16T17:23:33.000Z (about 1 year ago)
- Last Synced: 2025-01-16T18:44:05.081Z (about 1 year ago)
- Topics: computer-vision, dataannotations, datalabeling, linear-algebra, machine-learning, math, object-detection, opencv, python, yolo
- Language: Python
- Homepage:
- Size: 1.81 MB
- Stars: 1
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# augmentation
Twice the dataset, _same_ the time for annotation!
Multiply the size of the dataset (rotations, changes in brightness and contrast) in no time and don't waste another instant annotating it.
### Why?
Data augmentation is fundamental for a good object detection model. The problem is, the most boring part of object detection is annotating the dataset
(e.g. using https://pypi.org/project/labelImg/ for YOLO model), and having to do it multiple times because for slightly different images feels like
something that should be done automatically. And so it is (from now on)!
### How?
Put your annotated images in a directory with the annotation (.txt) files. The code assumes that for the image "my_image.jpg" there is the file "my_image.txt", structured as follows: `class_id center_x center_y bounding_box_width bounding_box_height` (see examples in [resources/](resources/)), where `center_x` and `bounding_bow_width` are normalized to the image width and `center_y`, `bounding_box_height` to the image height.
Now:
```
git clone https://github.com/davide710/augment-it.git
cd augment-it
pip install numpy opencv-contrib-python
```
You can see a demo of how the code works by running `python3 demo.py`. This shows how the images (data I used for an actual object detection model) are rotated and how the bounding box automatically follows the object thanks to the magic of linear algebra.
To use the functions in your code, simply
```
from rotation import rotate
from brightness_contrast import change_brightness_contrast
rotate(file_list, min_angle=5, max_angle=355, make_annotations=True, rotate_bbox=False, debug=False)
change_brightness_contrast(file_list, min_alpha=0.5, max_alpha=1.5, min_beta=-60, max_beta=60, make_annotations=True)
```
+ use `rotate_bbox=False` if your objects are circular (hence you just need to rotate the center)
+ use `make_annotations=False` if you just want the new images but not the new annotation files
+ `rotate` will save the output files in the directory "rotated/" and add "_rot" to all file names
+ `change_brightness_contrast` will save the output files in the directory "br_and_con/" and add "_bc" to all file names
#
#