‎

Paper acceptance at GCPR 2023

09.09.2023|15:30 Uhr

Our paper „LMD: Light-weight Prediction Quality Estimation for Object Detection in Lidar Point Clouds“ by Tobias Riedlinger, Marius Schubert, Sarina Penquitt, Jan-Marcel Kezmann, Pascal Colling, Karsten Kahl, Lutz Roese-Koerner, Michael Arnold, Urs Zimmermann and Matthias Rottmann has been accepted for publication and oral presentation at GCPR 2023.

Object detection on Lidar point cloud data is a promising technology for autonomous driving and robotics that has seen a significant increase in performance and accuracy in recent years. In particular, uncertainty estimation is a critical component for downstream tasks, and deep neural networks remain error prone even for high confidence predictions. Previously proposed methods for quantifying prediction uncertainty tend to modify the training scheme of the detector or rely on prediction sampling, which results in significantly increased inference time. To address these two issues, we propose LidarMetaDetect (LMD), a lightweight post-processing scheme for prediction quality estimation. Our experiments show a significant increase in statistical reliability in separating true from false predictions. We propose and evaluate an additional application of our method that leads to the detection of annotation errors. Explicit sampling and a conservative number of annotation error suggestions indicate the viability of our method for large datasets such as KITTI and nuScenes. Our preprint is available at arxiv.org/pdf/2306.07835.pdf.