Detecting Organisms for Marine Video Surveillance

摘要

The way to better understand the marine life and ecosystems is to surveil and analyze the activities of marine life. Research on marine organisms is becoming increasingly popular because of the increased focus in recent years. In this paper, we design a novel framework, dubbed Efficient Marine Organism Detector (EMOD) for high-resolution marine video surveillance, to detect and monitor marine organisms in a realtime and fast fashion. Current state-of-the-art marine organism detectors are mainly based on computer vision techniques that make great progress in recent years, which essentially requires a relatively large amount of various data. The datasets used are from the National Oceanic and Atmospheric Administration (NOAA), including a total five annotated video datasets HabCam, MOUSS, AFSC DropCam, MBARI and NWFSC. Experiments are performed on these three datasets with current popular one-stage detection methods (RetinaNet and SSD) and two-stage detection methods (Faster R-CNN and Cascade R-CNN) in our marine detector respectively. Experimental results demonstrate that our framework is competitive and efficient.