Simulated Object Detection in Underwater Environments

This research examines if a computer vision model trained on two simulated environments can perform well when the objects are shown in either environment. The environments were chosen in the context of underwater navigation and are a pool and a murky pond. The data from these environments were trained for object classification using two object detection pipelines: tiered and one-shot. The tiered detector first classifies the environment, and then the object. The one-shot detector only classifies the object. Through our study we found there were no significant differences between the performance of the one-shot and the tiered object detection. With the precision and recall being within 1% for most of the classes.