Paper Review: ObjectAug: Object-level Data Augmentation for Semantic Image Segmentation

Paper link

Main image

The idea is quite simple but cool. Surprisingly, such an approach wasn’t published before.

The gist of it:

  • Take an image with a mask.
  • Detach a masked object.
  • Fill in the hole with image inpainting.
  • Augment the masked object and put it back onto the background.

Thanks to the fact that we have separate objects, we can apply different augmentations to different categories. And we can also augment the background image separately from our objects or add augmentation on the newly formed image.

The approach was tested on PASCAL VOC 2012, Cityscapes, and CRAG.

One of the interesting things they checked was the performance of the approach for rare and hard categories.

The comparison

And there are a lot of ablation studies.

paperreview deeplearning augmentation imageinpainting imagesegmentation