Synthetic images of corals (Desmophyllum pertusum) with object detection models

Two object detection models using Darknet/YOLOv4 were trained on images of the coral Desmophyllum pertusum from the Kosterhavet National Park. In one of the models, the training image data was amplified using StyleGAN2 generative modeling.
The dataset contains 2266 synthetic images with labels and 409 original images of corals used for training the ML model. Included is also the YOLOv4 models and the StyleGAN2 network.

The still images were extracted from raw video data collected using a remotely operated underwater vehicle.
409 JPEG images from the raw video data are provided in 720x576 resolution. In certain images, coordinates visible in the OSD have been cropped.
The synthetic images are PNG files in 512x512 resolution.
The StyleGAN2 network is included as a serialized pickle file (*.pkl).
The object detection models are provided in the .weights format used by the Darknet/YOLOv4 package. Two files are included (trained on original images only, trained on original + synthetic images).

The machine learning software packages used is currently (2022) available on Github:
StyleGAN2: https://github.com/NVlabs/stylegan2
YOLOv4: https://github.com/AlexeyAB/darknet

本数据集基于来自科斯特哈维特国家公园的杯形珊瑚（Desmophyllum pertusum）图像，训练了两款搭载Darknet/YOLOv4框架的目标检测模型。其中一款模型的训练图像数据通过StyleGAN2生成式建模完成了数据增强。

本数据集包含2266张带标注的合成图像，以及409张用于训练机器学习（Machine Learning, ML）模型的原始珊瑚图像，同时提供了训练得到的YOLOv4模型与StyleGAN2网络。

本次所用静态图像均提取自遥控水下机器人采集的原始视频数据，其中从该原始视频数据中提取的409张JPEG图像分辨率为720×576，部分图像内屏幕显示（On-Screen Display, OSD）自带的坐标信息已被裁剪。

合成图像为分辨率512×512的PNG格式文件。

StyleGAN2网络以序列化pickle文件（*.pkl）的形式提供。

两款目标检测模型以Darknet/YOLOv4套件所采用的.weights权重格式封装，共包含两个权重文件：分别为仅基于原始图像训练的模型，以及基于原始图像与合成图像联合训练的模型。

本数据集所使用的机器学习软件包目前（2022年）可在GitHub平台获取：
StyleGAN2：https://github.com/NVlabs/stylegan2
YOLOv4：https://github.com/AlexeyAB/darknet