2021/04/22 by Alejandro Barrera, Barrera, Alejandro, Jorge Beltrán +7 · 2 citations
Computer Science · #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences
paper · pdf · doi:10.48550/arxiv.2104.11021
openalex publication_date 2021/04/22 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
The performance of object detection methods based on LiDAR information is\nheavily impacted by the availability of training data, usually limited to\ncertain laser devices. As a result, the use of synthetic data is becoming\npopular when training neural network models, as both sensor specifications and\ndriving scenarios can be generated ad-hoc. However, bridging the gap between\nvirtual and real environments is still an open challenge, as current simulators\ncannot completely mimic real LiDAR operation. To tackle this issue, domain\nadaptation strategies are usually applied, obtaining remarkable results on\nvehicle detection when applied to range view (RV) and bird's eye view (BEV)\nprojections while failing for smaller road agents. In this paper, we present a\nBEV domain adaptation method based on CycleGAN that uses prior semantic\nclassification in order to preserve the information of small objects of\ninterest during the domain adaptation process. The quality of the generated\nBEVs has been evaluated using a state-of-the-art 3D object detection framework\nat KITTI 3D Object Detection Benchmark. The obtained results show the\nadvantages of the proposed method over the existing alternatives.\n