This study focuses on evaluating the performance of object detection models, including Faster R-CNN, RetinaNet, YOLOv8, and Single Shot Detector (SSD), all optimized with Detectron2, for real-time detection of aquatic wildlife, particularly fish, in underwater environments. Marine research facilities worldwide collect extensive underwater image and video data, necessitating species-level classification of aquatic life. Traditional manual identification methods are time-consuming and prone to errors. This research aims to automate the real-time detection process, reducing the dependency on ecology experts and significantly expediting insights for researchers. The models were evaluated on underwater footage from the South African Institute of Aquatic Biodiversity, as well as the DeepFish and Fish4Knowledge datasets. The results indicate that YOLOv8 exhibits the fastest performance, achieving 31 to 68 frames per second (FPS), while maintaining acceptable accuracy, making it suitable for real-time fish detection. In contrast, Faster R-CNN and RetinaNet perform significantly slower, struggling to process some datasets efficiently. The findings contribute to the field of underwater fish detection by highlighting the optimal model and dataset configurations for achieving real-time performance.