Data collection and scene understanding have become crucial tasks in the development of intelligent vehicles, particularly in the context of autonomous driving. Deep Learning (DL) and transformer-based architectures have emerged as the preferred methods for object detection and segmentation tasks. However, DL-based methods often require extensive training with diverse data, posing challenges in terms of data availability and labeling. To address this problem, techniques such as transfer learning and data augmentation have been adopted. Simulators like CARLA have gained popularity in the autonomous driving domain, enabling the evaluation of architectures in realistic environments before real-world deployment. Synthetic data generated by simulators offers several advantages, including cost-effectiveness, access to diverse scenarios, and the ability to generate accurate ground truth annotations. In this paper, we focus our investigation on evaluating the performance and domain adaptation capabilities of a 3D object detection pipeline based on depth estimation using a stereo camera in the CARLA simulator. Our main objective is to analyze the results of the depth estimation stage using two different approaches: CoEx and SDN. The different experiments will be performed on real and synthetic scenarios from the KITTI and SHIFT datasets.
Exploring Domain Adaptation with Depth-Based 3D Object Detection in CARLA Simulator
Lect. Notes in Networks, Syst.
Iberian Robotics conference ; 2023 ; Coimbra, Portugal November 22, 2023 - November 24, 2023
27.04.2024
13 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Towards LiDAR and RADAR Fusion for Object Detection and Multi-object Tracking in CARLA Simulator
Springer Verlag | 2022
|DOAJ | 2023
|CARLA Simulated Data for Rare Road Object Detection
IEEE | 2021
|