Google's HDR+ image processing is almost magical at times, and it's all thanks to machine learning. Just a few years ago, Google's cameras were below average, but the Pixel changed all that with refined image processing algorithms. Now, Google is giving other research teams a chance to come up with their own version of HDR+ and compare it to Google's. It's giving away thousands of images in a gigantic data set, and you can check it out now.
HDR+ images are a composite of several full-resolution burst shots. Google's processing technology knows how to merge them in such a way that you get better exposure in both light and dark regions in the final image. You need a huge data set to develop such a model, and Google has provided it in the form of 3,640 bursts of full-resolution raw images, totaling 28,461 individual images. The archive is available under the Creative Commons license, and it isn't one of the "non-commercial" versions. Theoretically, you could build a commercial product using this dataset.
Google has already employed the dataset in a research paper, which uses a neural network to recreate part of the HDR+ pipeline. Several more papers using the dataset are on the way, too. If you want to check out Google's HDR+ dataset, it's available for browsing (with a Google account) on Google Cloud. You can also download the entire thing directly, but it weighs in at a hefty 765GB. All the files are DNG format.