The new dataset contains segmentation masks for 2.8 million object instances within 350 categories. The segmentation masks were produced with Google’s interactive segmentation process where human annotations work in close collaboration with AI models for segmentation in order to correct the neural network’s predictions.
This kind of collaboration was introduced by researchers at Google to improve the accuracy of the annotated images and at the same time speed up the process of annotation or segmentation mask drawing. The process of full manual annotation is time-consuming and requires a lot of effort from professional annotators.
Open Images v5 features around 100K segmentation masks for the validation and test sets with higher quality annotations. These masks are supposed to capture very fine details of complex objects and challenge machine learning algorithms in the task of image segmentation.
The updated dataset includes 6.4 million new image-level labels with around 20 000 categories. Together with the dataset, Google released the second Open Images Challenge which will include a new track for instance segmentation based on the improved Open Images Dataset.