TY - JOUR A1 - Tabata, Alan Naoto A1 - Zimmer, Alessandro A1 - Coelho, Leandro Dos Santos A1 - Mariani, Viviana Cocco T1 - Analyzing CARLA ’s performance for 2D object detection and monocular depth estimation based on deep learning approaches JF - Expert Systems with Applications UR - https://doi.org/10.1016/j.eswa.2023.120200 Y1 - 2023 UR - https://doi.org/10.1016/j.eswa.2023.120200 SN - 0957-4174 VL - 2023 IS - 227 PB - Elsevier CY - Amsterdam ER - TY - CHAP A1 - Lopes da Silva, Joed A1 - Tabata, Alan Naoto A1 - Cardoso Broto, Lucas A1 - Pereira Cocron, Marta A1 - Zimmer, Alessandro A1 - Brandmeier, Thomas ED - Campilho, Aurélio ED - Karray, Fakhri ED - Wang, Zhou T1 - Open Source Multipurpose Multimedia Annotation Tool T2 - Image Analysis and Recognition: 17th International Conference, ICIAR 2020, Proceedings, Part I UR - https://doi.org/10.1007/978-3-030-50347-5_31 KW - open source KW - video annotation KW - multipurpose KW - generic annotation Y1 - 2020 UR - https://doi.org/10.1007/978-3-030-50347-5_31 SN - 978-3-030-50346-8 SN - 978-3-030-50347-5 SN - 0302-9743 SN - 1611-3349 SP - 356 EP - 367 PB - Springer CY - Cham ER - TY - THES A1 - Tabata, Alan Naoto T1 - Object detection and monocular depth sstimation with a custom synthetic automotive dataset N2 - In the automotive industry, the concept of autonomous vehicles is becoming closer to reality, with companies disputing to be the pioneers on reaching at least a level 3 on driving automation. However, before implementing autonomous vehicles on a large scale, research and testing should be performed to assess its safety and reliability. Since one of the ways autonomous vehicles sense its surrounding is through cameras, then one approach to promote human safety is by researching computer vision techniques that may help the vehicle to better understand the context it is in. Therefore, on this work algorithms capable of detecting pedestrians and vehicles, and their distance to the camera are evaluated, in a way that future works can apply corrective trajectory procedures in advance. The main contributions of this work are application and validation of such techniques in a context different from those of which have already been extensively tested on the literature. In this dissertation, this is done by creating a custom CARLA-based synthetic dataset and evaluating its knowledge transfer capability with computer vision algorithms to a real-world dataset, Waymo Open. The purpose of a synthetic dataset is the possibility of generating huge amounts of data at will, a requirement for parametrizing state-of-the-art computer vision models based on deep convolutional neural networks. The Faster R-CNN with a ResNet 50 as backbone was evaluated for the bounding box task, and for monocular depth estimation, the monodepth2 model with a U-Net and ResNet 18 as backbone was evaluated. On the object detection part, it was noted that the injection of synthetic data did not aid in model generalization, with 12% performance decrease when compared to training from scratch on the Waymo skip 10 dataset. For monocular depth estimation, however, the best performing models proved to be different combinations of both synthetic and real-world data, with them improving the performance metrics on average 5% on the Waymo dataset. Overall, it is noted the importance of data variety for both algorithms, with the current synthetic dataset iteration being beneficial for monodepth2 but not Faster R-CNN, which suggests that there is still room for improvement. These observations lead to the conclusion that features which impact positively the model for creating a dataset differ according to the algorithm’s purpose, and as such the creation of an all-purpose dataset is probably not ideal. Y1 - 2020 UR - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:573-8625 CY - Ingolstadt ER -