2021/08/18 by Lukas Stäcker, Juncong Fei, Stäcker, Lukas +11
Computer Science · Engineering · #Advanced Neural Network Applications #CCD and CMOS Imaging Sensors #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Industrial Vision Systems and Defect Detection #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2108.08166
openalex publication_date 2021/08/18 · openalex created_date 2025/10/24 · openalex updated_date 2026/07/28
Deep neural networks have proven increasingly important for automotive scene\nunderstanding with new algorithms offering constant improvements of the\ndetection performance. However, there is little emphasis on experiences and\nneeds for deployment in embedded environments. We therefore perform a case\nstudy of the deployment of two representative object detection networks on an\nedge AI platform. In particular, we consider RetinaNet for image-based 2D\nobject detection and PointPillars for LiDAR-based 3D object detection. We\ndescribe the modifications necessary to convert the algorithms from a PyTorch\ntraining environment to the deployment environment taking into account the\navailable tools. We evaluate the runtime of the deployed DNN using two\ndifferent libraries, TensorRT and TorchScript. In our experiments, we observe\nslight advantages of TensorRT for convolutional layers and TorchScript for\nfully connected layers. We also study the trade-off between runtime and\nperformance, when selecting an optimized setup for deployment, and observe that\nquantization significantly reduces the runtime while having only little impact\non the detection performance.\n