Our Process

We first print vinyl posters of the competition objects and reconstruct what we anticipate to be the obstacles. Then, we submerge the objects underwater and take videos. From the underwater videos, we use a Python and OpenCV to publish and extract frames which will be the dataset to be trained on. After we have prepared our dataset, we then manually annotate each image with bounding boxes. Once annotations are done, the dataset is ready to be trained on. We employ a few augmentations to make our dataset more robust and help improve accuracy.

Additionally, for some obstacles, we may add classes to denote orientation which helps Autonomy make decisions on navigation. We then use Google Colab for its computing ability (GPU) to train using the YOLO algorithm. Once we have our trained object detection models, we utilize OpenCV within our submarine to subscribe and publish frames from the equipped RGB camera using ROS.

Before Applying Augmentation (Static) -

After Applying Augmentation (Static) -