eduzhai > Applied Sciences > Engineering >

Cityscapes 3D Dataset and Benchmark for 9 DoF Vehicle Detection

  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Detecting vehicles and representing their position and orientation in thethree dimensional space is a key technology for autonomous driving. Recently,methods for 3D vehicle detection solely based on monocular RGB images gainedpopularity. In order to facilitate this task as well as to compare and drivestate-of-the-art methods, several new datasets and benchmarks have beenpublished. Ground truth annotations of vehicles are usually obtained usinglidar point clouds, which often induces errors due to imperfect calibration orsynchronization between both sensors. To this end, we propose Cityscapes 3D,extending the original Cityscapes dataset with 3D bounding box annotations forall types of vehicles. In contrast to existing datasets, our 3D annotationswere labeled using stereo RGB images only and capture all nine degrees offreedom. This leads to a pixel-accurate reprojection in the RGB image and ahigher range of annotations compared to lidar-based approaches. In order toease multitask learning, we provide a pairing of 2D instance segments with 3Dbounding boxes. In addition, we complement the Cityscapes benchmark suite with3D vehicle detection based on the new annotations as well as metrics presentedin this work. Dataset and benchmark are available online.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×