Blue mussel reefs are listed as biogenic reefs in the Habitat Directive of the European Union. They can protect shorelines and function as habitats for many marine creatures, but they are in danger. To protect existing mussel reefs, it is necessary first to find mussels and then map them to confirm that they form a mussel reef. DTU Aqua uses ROV(remotely operated vehicle) to film seabed and map mussel reefs based on ROV videos. We propose one method based on YOLO(You Only Look Once) algorithm to detect mussels in the videos, roughly estimate their coverage automatically, and indicate potential mussel reefs. Additionally, one dataset, including 1000 labeled mussel images, was constructed to bridge the gap in marine datasets.
03-02_16.07.53: The name is the video's name. It contains all the .npy files corresponding to all mussels' coverage in all frames in the video.
yolov5_modified: It includes all the code and trained model: best.pt.
yolov5_modified/segment: It includes all code used for image segmentation.
Poster.pdf and Report.pdf are the poster and report to this project.
test_images: It contains three images can be used as source input to detect.py for mussel detection.
example.gif: It is the clip of the output video after detection. The complete video is very big around 8G nearly thirty minutes.
Firstly, download the YOLO code yolov5_modified, and use detect.py to detect and mark mussels in videos with rectangles. Same as the turtorial in the https://github.com/ultralytics/yolov5/blob/master/tutorial.ipynb. For example, "python detect.py --weights best.pt(trained model) --source xxxxx(path of video)".
Secondly, one file named by the video's name is created, it includes .npy files archived can be check later, the number of them is the number of frames in the video. And one graph(time vs mean coverage) will be generated.
For segmentation, it's the same as the turtorial in https://github.com/ultralytics/yolov5/blob/master/segment/tutorial.ipynb. In directory yolov5_modified/segment, run command "python predict.py --data musselreef_seg_4.yaml --weights best.pt(trained model) --source xxxxx(path of video)" to mark mussels in videos with red masks and bounding boxes. Make sure correct weights and configuration for image segmentation is used.