Static Traversability Dataset
The static traversability dataset consists of 100 images of indoor/outdoor scenes picked from the Internet. These images were deliberately chosen to be challenging, with office chairs, highly reflective surfaces, specular reflections, traversable surfaces for which multiple ground models would be necessary, and environments in which the traversable surface has very similar if not indistinguishable appearance to obstacle regions.
Testing a traversability detection algorithm on static images has many advantages. It allows the performance of the algorithm to be evaluated in multiple environments quickly.
Moreover, the performance of the algorithm on static images from a multitude of environments is tough since the traversability detection algorithm must generalize over many variations in appearance, illumination etc..
From static images, pixels available for a traversability model are limited and cannot be collected over time. Since no ground truth information was readily available for these images, a human observer was asked to manually label the traversable image regions. All images were rescaled to 160x120.