|
The OmniScape dataset contains, for each capture, fisheye and catadioptric stereo RGB images from the two
front sides of a motorcycle, with semantic segmentation
and depth map ground truth, as well as the dynamics of
the vehicle with its velocity, angular velocity, acceleration
and orientation. Currently, the OmniScape dataset contains more than 10,000 captures, and
will be progressively augmented with more omnidirectional
data using different vehicles,
modalities and environments. The dataset contains data generated from GTA V and CARLA, and can be extended
to other simulators. The RGB images are
available for 14 different weather conditions and time of
the day and this for each capture. CARLA simulator gives a semantic segmentation into 13 classes, namely Building, Fence, Other,
Pedestrian, Pole, Road line, Road, Sidewalk, Vegetation,
Vehicle, Wall, Traffic sign, Unlabeled. In complement to these omnidirectional images, the OmniScape dataset
contains also the dynamics of the vehicle at
each capture, such as velocity, angular velocity, acceleration
and orientation. The case of twowheelers is more challenging because of the dynamics of
these vehicles. These alterations will for sure affect classical tasks such as
visual odometry and semantic segmentation. This is due to
the fact that most computer vision and machine learning tasks
are often trained on data acquired with cars as autonomous
vehicles, while these vehicles do not suffer from modifications in these dynamics.
|
