PandaSet+: Enhanced and Augmented Annotations for Panoptic Perception
Zhe Peng, Lei Qin, Cheng Zhang, Chi Yun Xu, Lingyu Wang, Rachid Benmokhtar, Xavier Perrotton · 2024
PandaSet is a valuable dataset for autonomous driving perception. It provides dense and long-range LiDAR point clouds, synchronized with 6 perspective cameras with different orientations. However, recorded and released by a leading LiDAR vendor, the PandaSet is strongly oriented to point-cloud processing, making its usage for computer vision applications less convenient. In this paper, we present PandaSet+, an enhanced and augmented version of the existing PandaSet dataset. Firstly, we enhanced the 3D object annotations by filtering out objects that are severely occluded by road infrastructures or by other objects. Secondly, we augmented the dataset with 2D and 3D lane annotations. Further, we transformed all the 3D annotations from the global world coordinate system to the coordinate system of each camera. The resulting PandaSet+ is suitable for panoptic perception, e.g. 3D objects detection, 2D and 3D lanes detection from images. Finally, we trained and evaluated state-of-the-art algorithms on PandaSet+, therefore validated its usefulness and established baseline performances. Like PandaSet, PandaSet+ will be released under the CC BY 4.0 license that supports free commercial use. The Project is available at https://github.com/leiqin1/pandaset_plus.