The need for dynamic and active datasets

Abstract

Datasets provide open and accessible benchmarks for the refinement of existing algorithms and the testing of new techniques. Notably, recent advances in computer vision, particularly in convolutional and deep neural networks, have been enabled by the availability of appropriate datasets. Though datasets have evolved to contain ever more data, the focus has remained on static and passive information. We propose that in emerging application areas such as augmented reality and robotics, a more active and dynamic approach to datasets is required. We review key limitations of existing datasets, and propose directions for discussion and creation of deeper dataset capabilities, including accommodating shifts in camera pose, scene behavior and affordances, and variations in camera configuration