This workshop focuses on the following important topics in the context of Localization, SLAM, and VO:
- Common to existing approaches to the Visual Localization problem, whether they rely on local features or CNNs, is that they generate a representation of the scene from a set of training images. These approaches (implicitly) assume that the set of training images covers all relevant viewing conditions. In practice, this assumption is typically violated as it is nearly impossible to cover complex scenes over the full range of viewing conditions. Moreover, many scenes are dynamic: the geometry and appearance of scenes changes significantly over time, e.g., due to seasonal changes in outdoor scenes or changes in furniture in indoor scenes. This workshop aims to serve as a benchmark for the current state of visual localization under changing conditions and to encourage new work on this challenging problem.
- We have see impressive progress on Visual SLAM (V-SLAM) with both geometric-based methods and learning-based methods. However, none of those methods is robust enough for high-reliability robotics, where challenging situations such as changing or a lack of illumination, dynamic objects, and texture-less scenes, exist and no other sources of odometry are available. Unfortunately, popular benchmarks such as KITTI or TUM RGB-D SLAM are too clean and simple, have rather restricted motion patterns, usually only cover one type of scene (e.g. urban street, indoor), and are often free of degrading effects such as lighting changes and motion blur. This workshop puts forth a challenge to gather evidence on the robustness of geometric and learning-based SLAM in challenging situations and to push the limit of geometric and learning-based SLAM towards real world applications. To this end, the workshop provides a new benchmark with large high-quality and diverse data and good labels.
- The development of smart-phones and cameras is also making the visual odometry more accessible to common users in daily life. With the increasing efforts devoted to accurately computing the position information, emerging applications based on location context, such as scene understanding, city navigation and tourist recommendation, have gained significant growth. The location information can bring a rich context to facilitate a large number of challenging problems, such as landmark and traffic sign recognition under various weather and light conditions, and computer vision applications on entertainment based on location information, such as Pokemon. This workshop solicits scalable algorithms and systems for addressing the ever increasing demands of accurate and real-time visual odometry, as well as the methods and applications based on the location clues.
Besides offering concrete challenges, invited talks by experts from both academia and industry provide a detailed understanding about the current state of Visual Localization, SLAM, and VO algorithms, as well as open problems and current challenges. In addition, the workshop solicits original paper submissions.