SUNOD: Synthetic Underwater Non-Natural Object Detection Dataset

Deep-learning-based non-natural object detection in complex underwater environments is widely needed in many application scenarios. However, obtaining image data of underwater non-natural objects is difficult, so there is basically no relevant dataset. Besides, the underwater environments and lighting conditions are complex, which will lead to image quality degradation and increase the difficulty of detecting. For this purpose, we build a Synthetic Underwater Non-natural Object Detection (SUNOD) dataset. We first utilize a free software, Unreal Engine 5 (UE5), to create a realistic three-dimensional virtual underwater environment containing 12 kinds of non-natural objects. Then we take pictures of these non-natural objects and manually label 2,840 images. To make SUNOD more challenging and be able to create object detection models with more generalization capabilities and better robustness, we expand it to contain 72,943 images by 26 kinds of data augmentation methods. In addition, we report the performance of 5 YOLOv8 object detection models with different volumes on SUNOD. SUNOD can be used both to fine-tune pretrained models to accomplish the underwater non-natural object detection task and to be a benchmark to evaluate new models for this task. This study provides the first synthetic dataset for underwater non-natural object detection, which contributes to the research on the intelligence of submarines, exploration robots, and other underwater devices.