Burst Image Deblurring Using Permutation Invariant Convolutional Neural Networks
Miika Aittala and Fredo Durand

This archive contains a supplemental appendix document, and result images on
datasets from previous work, as well as our new datasets.

Due to file size constraints, we only include results on full bursts in this
package (without the progressions shown in the paper), and heavily compress
all images.  Hence, these images are provided for visual reference only and
should not be used for evaluating algorithms. We will release the uncompressed
data, results, source code and pre-trained weights for the method separately
to facilitate future research. 

Datasets we used from Delbracio et al. (2015) and Kohler et al. (2012), as
well as result images from methods prior to Wieschollek et al. (2017) can be
found at:

http://dev.ipol.im/~mdelbra/fba/
https://github.com/cgtuebingen/learning-blind-motion-deblurring/releases
http://webdav.is.mpg.de/pixel/benchmark4camerashake/src_files/Instructions_authors.html

The new datasets we introduce are 'books', 'car_snow', 'dishwasher', 'street' 
and 'utensils'.

The files in the folders are as follows:

- inputs/: the input images for the burst (for our new datasets only)
- OURS.jpg: our result when fed with all images of the burst
- rdn.jpg: result of Wiescholleck et al. (2017) on the same input, computed 
  using their software implementation from the address above, or provided by
  the authors

books dataset also contains a validation photo, which can be used to verify
that the recovered text is accurate, to the extent that it is readable.