Autonomous driving has recently emerged as a keystone problem for computer vision and machine learning, with significant interest in both academia and industry. Besides being a rich source of research problems for visual perception, learning, mapping and planning, it is also poised to have an immense societal and economic impact. Several large efforts from the automotive industry have projected the imminent deployment of Level 3 autonomous systems, with a few efforts also geared towards Level 5 autonomy in the near future. But reliable solutions have been trained and validated only in controlled environments, while a vast majority of road conditions deviate from the ideal. This workshop calls for intensive engagement from the research community to address this problem and proposes a benchmark dataset with rich annotations in relatively unconstrained conditions to facilitate the effort. A higher level goal is to percolate autonomous driving to domains where road infrastructure is sub-optimal for computer vision and machine learning but which stand to gain immeasurably from its benefits.
The theme of the workshop will be to explore feasibility and directions for the next generation of vision and learning solutions that will handle such real-world challenges. Consistent with this motivation, this edition introduces a domain adaptation challenge to study semantic segmentation across driving scenes from multiple domains, where labeled data might be limited in the target domains.
This workshop also adopts a broad view of what is entailed by driving in unconstrained environments. Some aspects such as changes in weather, time of day or imaging conditions are already being studied by the community, for instance, under the purview of domain adaptation. However, these have still largely been for environments with organized traffic and favorable infrastructure. Thus, besides those aspects, this workshop also poses the challenge of autonomous driving in less constrained traffic, along with the infrastructure that is not always dependable. In particular, we acquire our dataset on roads that present unique challenges for computer vision and machine learning, such as:
