Extending the Mapillary Vistas Dataset for Perfecting Street Scene Segmentation Models
Last year, we launched the Mapillary Vistas Dataset as the largest and most diverse publicly available street-level imagery dataset for teaching machines to “see”. The pixel-wise and instance-specifically annotated data enables training top-of-class semantic segmentation models. In the past year, our customers, such as Autonomous Intelligent Driving (the Audi subsidiary that builds the full software stack for autonomous driving across the Volkswagen group), as well as our own research team have used it to achieve outstanding results.
We’ve been happy to see more great street-level datasets published since then, such as the BDD 100K and ApolloScape with images from the US and China, respectively. However, when we look at the 350-million-image database on the Mapillary platform, covering street scenes all over the world, it’s clear what a unique opportunity it provides for getting diverse training data that would be very hard to obtain otherwise.
That is why we’ve worked out an extension of the Mapillary Vistas Dataset to further contribute to developing deep learning algorithms for mapping, GIS, and autonomous driving. Since its creation, the Mapillary Vistas Dataset is the most diverse street-level image dataset in terms of object classes. The extension adds even more granularity to the annotations, offering over 60 new classes and, as a completely new feature, traffic light states.
Sample annotations of the extended object classes (click on an image to view the large version)
More specifically, this is what’s included in the extension:
- Road markings: increased from 6 classes to 36 classes including speed limits and direction arrows, which are essential classes for building navigation and HD maps.
- Driveway: a new class that we’ve introduced to help enhance data extraction for last-mile delivery.
- Barriers: increased from 6 to 12 classes. Detailed refinement of the barrier classes is the first step to enhance the perception capacity for autonomous driving under different scenarios.
- Signages: the billboard class is now refined to advertisement, storefront, and general information signages. This facilitates automatic and in-depth location-aware data extraction for marketing, business, and point-of-interest analysis.
- Traffic light states: a new feature in the Mapillary Vistas Dataset where dynamic properties are associated with static objects. In this case, we have refined 100,000 traffic light annotations with different states including red, green, yellow, direction, and off. The diversity of the Vistas combined with state information will help further enhance traffic light recognition.
Annotations of traffic light states for different shapes of traffic lights in diverse scenes and lighting conditions
This update will further advance Mapillary's algorithms for extracting diverse map data from images, and we hope it will also speed up the developments in the industry and research. The extension is currently available for the commercial edition of the Mapillary Vistas Dataset, and we'll soon make some of the new classes available for the research edition as well. Please contact email@example.com for further details.
Note. The statistics on object classes have been updated in this post on 5 September 2018.