BboxCutMix

This is a data augmentation for object detection, using bounding boxes.

Introduction

There were no-objects images during training when using Mosaic[1] because of image center position of objects. So we implemented a augmentation pasting bboxes in the empty space of a training image(BboxCutMix).

An example of BboxCutMix

The distribution of bounding box sizes

Details

There is already a method using bounding boxes like copy-paste[2]. But BboxCutMix has two differences from copy-paste.

1. Make hard to find a bounding box when training using offset margin.
2. Have full edge information(no-occluded objects).

Data augmentation results

The detailed procedure of BboxCutMix is shown in the image below. The offset margin was added to prevent easy finding out a bounding box. And we applied data augmentations for each cropped objects. Finally, cropped objects are inserted into the training image.

The procedure of BboxCutMix

How to use

1. run the command

$ pip install -r requirements.txt

2. initialize variables and add the code in your Dataset class like below.

from bbox_cutmix import * 
from torch.utils.data import Dataset 

    class customDataset(Dataset):
        ...
        self.ims = ...        # training images (list | (#_img, H, W, C))
        self.im_files = ...   # path of traning images (list)
        self.img_size = ...   # traning images size (int)
        self.indices = ...    # range(0, #_img)
        self.labels = ...     # labels of traning data (list | (#_img, #_obj, class, x, y, w, h))
        …
        # load image
        if bbox_cutmix:
            # h, w are height and width of img
            img, labels = bbox_cutmix(self, img, labels, h, w)
        …

References

1. Zhiwei et al, AMRNet: Chips Augmentation in Aerial Images Object Detection, https://arxiv.org/abs/2009.07168
2. Golnaz et al, Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation, https://arxiv.org/abs/2012.07177

by seareale | KVL Lab | KNU