In this document, we show additional qualitative results to extend fig. 12, examples from our training dataset and additional results on partially a cloudy day.
In this section, we present additional qualitative results, extending fig. 12.
These conditions are typical sunny days in which PS is expected to be unstable or break, which explains the high amount of errors in the results. In comparison to prior work, our network models prior knowledge on those specific cases and thus outputs state of the art results.
In this section, we present a random subset of 100 groups of 8 input images from the training dataset.
Note that those images are not fed directly to our CNN. Non-empty patches of 16x16 pixels are extracted from those images for training.
In this section, we present results on the partially cloudy day shown below. We follow the same procedure as section 8.6 to render images on partially cloudy days taken from the Laval HDR sky dataset. Our method still shows competitive performance despite its degraded accuracy compared to clear weather.
Images are tonemapped using the gamma correction γ=2.2 for visualization.