Benchmarking edge computing devices for grape bunches and trunks detection using accelerated object detection single shot multibox deep learning models

Visual perception enables robots to perceive the environment. data is processed using computer vision algorithms that are usually time-expensive and require powerful devices process visual in real-time, which unfeasible for open-field with limited energy. This work benchmarks performance of different heterogeneous platforms object detection real-time. research three architectures: embedded GPU—Graphical Processing Units (such as NVIDIA Jetson Nano 2 GB 4 GB, TX2), TPU—Tensor Unit Coral Dev Board TPU), DPU—Deep Learning Processor AMD/Xilinx ZCU104 Development Board, Kria KV260 Starter Kit). The authors used RetinaNet ResNet-50 fine-tuned natural VineSet dataset. After trained model was converted compiled target-specific hardware formats improve execution efficiency. were assessed terms evaluation metrics efficiency (time inference). Graphical (GPUs) slowest devices, running at 3 FPS 5 FPS, Field Programmable Gate Arrays (FPGAs) fastest 14 25 FPS. Tensor (TPU) irrelevant similar TX2. TPU GPU most power-efficient, consuming about W. differences, metrics, across have an F1 70 % mean Average Precision (mAP) 60 %.