COCO CROPPER

Crops annotations (bounding boxes) in COCO dataset.

Configurables:

• Margin

  • with what margin the photo is going to be cropped

• Minimum area

  • minimum wanted area of the cropped photo in pixels (width*height)

• Minimum width

  • minimum wanted width of the cropped photo in pixels

• Minimum height

  • minimum wanted height of the cropped photo in pixels

• Classes

  • which classes are wanted among all dataset

Examples

• Below examples have been created with the following configuration:

  • margin = 0.12 (%)
  • min_area = 16384 (pixels)
  • min_width = 96 (pixels)
  • min_height = 96 (pixels)

Input 1

Output 1 - Vehicle

Input 2

Output 2 - Animal

Input 3

Output 3 - Person

Output 3 - Outdoor

Usage

0-) Download the COCO dataset with the annotations json from below link if you do not have them yet:

https://cocodataset.org/#download

1-) Clone this repository:

git clone https://github.com/senceryucel/coco-cropper

2-) Navigate into the folder:

cd coco-cropper

3-) Download prerequisites:

python3 -m pip install -r requirements.txt

4-) Run the script with the -h flag to see how to configure it:

python3 src/main.py -h

Example usage

python3 src/main.py -ij ~/path_to/instances_train2017.json -id ~/path_to/train2017 -o ~/path_to/outputs/ -m 0.12 -ma 16384 -mw 96 -mh 96 -c all

---

Sencer Yucel, 2023