Extract-subset-from-coco

In the past days, I have tried many methods to extract specific data from coco-format dataset.

Some dataset may not launched by coco officially, but any other dataset in coco-format, the dataset I used is FashionPedia, which contains many accessory imgs. I now share some experience and related code in the repo

The repo will attach 3 py files, or more, but generally could divided into 3 stage

So the code will be named Stage_{}, each file I will try to write some annotations.

Stage_1.py

In the py file, the code will use your own dataset to extract specific category and write a corresponding xml file for each img

Stage_2.py

In the stage, the code snippet will use the xml file generated in Stage_1 to generate a coco format json file.

If you use coco official dataset(coco_2017), the image name is a value from a certain number, so you could also named the variable <file_name> from the minimum value, and every loop the value will +1

Stage_3.py

If you only want get a coco format annotation, the first 2 stage is enough. In Stage_3, the code will also generate a YOLO format annotation. The txt file contain the following contents: <object-class> <x> <y> <width> <height>

where <object-class> is an integer representing the class of the object,

<x> and <y> are the normalized coordinates of the center of the bounding box,

and <width> and <height> are the normalized dimensions of the bounding box

verify.py

After generate a new data subset, you could also check if you make it in a correct format. For me, I use detectron2 to check its format. Before use it, install it, I also attach code in the py file, if your device is Windows, you may refer to the official repo.