Spatial Distribution Data Augmentations for Long Range LiDAR Object Detection
Publicerad
Författare
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Abstract
Since the KITTI dataset was introduced in 2013, research in LiDAR-based object detection (OD) has increased significantly in popularity. The spike in research has also led to improved performance, but the predictions of these models are still not robust or accurate enough to safely be used for autonomous driving. Although the overall performance of the models is insufficient, long-range predictions is an area
where the need for better accuracy is particularly apparent. The emergence of large datasets has improved the overall performance of such models, but most datasets still require data augmentations to help models converge and generalize well to unseen data. Many methods used for data augmentation in LiDAR OD have been accepted as standards in the field since they have been shown to improve overall
performance. Techniques like Ground Truth Database Sampling (GTDS) and Fade are used without thoroughly analyzing the class-specific effects and results they impose on the models. This thesis investigates how spatial distribution data augmentations can improve long-range accuracy of LiDAR OD. This is done by downsampling ground truth objects from the original GTDS and moving them further away from the sensor. The goal is to increase diversity in the long-range data while keeping a realistic spatial distribution of points for objects sampled at that range. Furthermore, a classspecific analysis of the widely accepted GTDS and Fade augmentations is conducted to further explore the effects these data augmentations have on the popular nuScenes dataset. All experiments are conducted on the two common benchmarking models PointPillars and CenterPoint. The work shows that GTDS can negatively impact the detection accuracy of less frequent classes in nuScenes, even if the overall accuracy increases. Moreover, we demonstrate the importance of using Fade in conjunction with GTDS and how it can mitigate class-specific accuracy losses introduced by GTDS. Lastly, this thesis analyzes why the long-range sampler ultimately fails at increasing long-range accuracy.