https://github.com/yanndubs/overlayed-datasets

Code for generating datasets consisting of overlayed images.
https://github.com/yanndubs/overlayed-datasets

classification computer-vision datasets occlusion overlay python

Last synced: 7 months ago
JSON representation

Code for generating datasets consisting of overlayed images.

• Host: GitHub
• URL: https://github.com/yanndubs/overlayed-datasets
• Owner: YannDubs
• License: mit
• Created: 2019-10-04T04:43:36.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2019-10-07T18:46:55.000Z (almost 6 years ago)
• Last Synced: 2025-01-24T09:11:23.558Z (9 months ago)
• Topics: classification, computer-vision, datasets, occlusion, overlay, python
• Language: Jupyter Notebook
• Homepage:
• Size: 477 KB
• Stars: 1
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Overlayed-Datasets

Datasets consisting of overlayed images. The idea of overlaying datasets came after reading [Information Dropout](https://arxiv.org/abs/1611.01353).

## Use Precomputed datasets

2 precomputed datasets can be found in `data/` (the images, labels, and distracting - overlayed - labels), namely `CIFAR10 + MNIST` and `MNIST + OMNIGLOT`. Here are some examples for both of these:

![CIFAR10 + MNIST](imgs/cifar10_mnist.png)

![MNIST + OMNIGLOT](imgs/mnist_omniglot.png)

The first label is the real (background), while the second corresponds to the distractor (overlayed) label.

The code to generate those, and examples with additional arguments can be found [examples.ipynb](/examples.ipynb).

## Generate Your Own

To generate your own, you need `overlay_img` (or the wrapper `overlay_save_datasets` if you directly want to save it).

The main documentation is in the following docstring:

```python

def overlay_img(bckgrnd, to_overlay, is_shift=False, seed=123):

    """Overlays an image with black background `to_overlay` on a `bckgrnd`

    

    Parameters

    ----------

    bckgrnd : np.array, shape=[n_bckgrnd, height_bckgrnd, width_bckgrnd, ...], dtype=uint8

        Background images. Each image will have one random image from  `to_overlay` overlayed on it.

    

    to_overlay : np.array, shape=[n_overlay, height_overlay, width_overlay, ...], dtype=uint8

        Images to overlay. Currently the following assumptions are made:

            - the overlaid images have to be at most as big as the background ones (i.e. 

              `height_bckgrnd <= height_bckgrnd` and `<= width_bckgrnd`).

            - The overlayed images are also used as mask. This is especially good for black 

              and white images : whiter pixels (~1) are the ones to be overlayed. In the case

              of colored image, this still hold but channel wise.

    is_shift : bool, optional

        Whether to randomly shift all overlayed images or to keep them on the bottom right.

        

    seed : int, optional

        Pseudo random seed.

        

    Return

    ------

    imgs : np.array, shape=[n_bckgrnd, height, width, 3], dtype=uint8

        Overlayed images.

        

    selected : np.array, shape=[n_bckgrnd], dtype=int64

        Indices of the slected overlayed images.

    """

```