Project link: https://github.com/openimages/dataset

Open Images is a new dataset from Google, with over nine million URLs spanning over 6000 categories. Each image is processed by Google's vision model and verified by a human. As of July 20, 2017, there are also over two million bounding box annotations spanning over 600 objects.

The difference is that Open Images covers more real-life objects than others, which can be very useful when developing real-life applications.