Edge AI is redefining the deployment of computer vision systems by enabling real-time inference directly on resource-constrained edge devices. This shift offers significant advantages in terms of reduced latency, data privacy, and operational autonomy in bandwidth-limited computing environments. In this paper, we present a systematic performance benchmarking of multiple variants of YOLOv8 on the NVIDIA Jetson Orin NX platform, focusing on object detection tasks. We evaluate inference latency, frame throughput, and computational resource usage across varying input sizes and model complexities. Furthermore, we validate the deployment effectiveness through practical use cases, such as vehicle and package detection. Our findings show that the TensorRT model outperforms PyTorch by 17.7% at a batch size of 2, although PyTorch presents greater stability at larger batch sizes (e.g., 8), where TensorRT encounters resource constraints. In terms of memory usage, it increases linearly with batch size: 69% from batch 1 to 4, with TensorRT requiring 429.20 MB at batch size 2 compared to PyTorch’s 451.24 MB. Furthermore, the processing time per image decreases by 42% when scaling from batch size 1 to 4, highlighting a critical saturation point for edge resources. In summary, the results provide insight into the trade-offs between model size and speed, offering guidance for selecting detection architectures tailored to real-time edge applications.

Underwater object detection faces significant challenges including uneven lighting, low contrast, and scattering-induced distortions. Existing underwater datasets are limited in scale, class diversity, and annotation consistency, which hinders robust model development. This work addresses these limitations by creating the UWOD dataset through the integration of seven publicly available underwater datasets, comprising over 107 K images with approximately 374 K annotations across 39 classes. We employ a semi-automatic annotation pipeline that combines manual labeling with iterative model-in-the-loop training to ensure high-quality ground truth, As a final verification step, all auto-generated labels were manually inspected and corrected as needed. We benchmark state-of-the-art object detectors—including YOLOv8, YOLOv7, YOLOv5, FCOS, EfficientDet, YOLOX, RT-DETR, SSD, and Faster R-CNN—establishing comprehensive performance baselines; YOLOv8 and YOLOv7 achieve the best accuracy–efficiency trade-off. Our transfer learning analysis shows that domain-specific pretraining substantially often outperforms pretraining on general-purpose datasets, yielding up to more than 50% improvement in low-data regimes versus training from scratch, with markedly lower seed-to-seed variance than scratch training. Sequential pretraining on COCO followed by UWOD achieves the strongest results on our most challenging dataset. Due to upstream licensing constraints, we release trained model weights and an automated annotation pipeline that encapsulate the learned underwater-domain knowledge, enabling immediate application to new imagery while respecting intellectual property.