BENCHMARKING 2D AND 3D OBJECT DETECTION MODELS ON DIFFERENT EDGE COMPUTING PLATFORMS

The deployment of artificial intelligence models on edge computing platforms is imperative for applications necessitating real-time perception, including object detection and gesture recognition. This paper presents a thoroughgoing benchmarking study of 2D and 3D convolutional neural network models across a range of representative edge devices, including the NVIDIA Jetson AGX Orin, Jetson Nano, and Raspberry Pi 3. The study utilizes gesture recognition and mask detection as the primary use cases for evaluating the performance of these models. We evaluated the inference time, memory footprint, and power consumption under varying input resolutions and sequence lengths, thereby capturing the latency–accuracy–energy trade-offs that govern practical deployment. The results show that AGX Orin consistently handles high-demand scenarios with low latency and stable throughput, whereas Jetson Nano offers competitive efficiency in energy-constrained settings. Raspberry Pi 3, while limited, establishes a useful baseline for CPU-only operation and highlights optimization needs. We also examined the effects of quantization and runtime optimizations and reported configurations that deliver robust performance without substantial accuracy loss. This study synthesizes these findings into actionable guidelines for selecting models, precisions, and device classes according to operational constraints. By providing a reproducible evaluation framework and device-aware recommendations, this study supports developers in balancing speed, accuracy, and energy efficiency, thereby enabling reliable real-time AI on resource-limited edge platforms. Our methodology uses warm-up iterations, repeated trials per configuration, synchronized power sampling, and pre-processing to ensure fair comparisons.