Seeing the Unseen: Adaptive-EIoU (A-EIoU) for Robust Underwater Object Detection
Nipun Taneja
Department of Computer Science
Faculty Supervisor: Akila De Silva
Monitoring marine ecosystems is vital for climate tracking, but automating this with AI is difficult because standard object detection metrics strictly assume rigid, geometric shapes. These metrics arbitrarily penalize the amorphous, irregular boundaries of marine invertebrates like starfish, sea urchins (Echinus), and sea cucumbers (Holothuroidea). To address this, we introduce Adaptive-EIoU (A-EIoU), a novel bounding box regression loss function integrated into the YOLOv8-nano architecture.
A-EIoU utilizes a tunable "rigidity" parameter that relaxes strict aspect-ratio constraints, allowing the neural network to mathematically accommodate biological flexibility. Tested on the Detection of Underwater Objects (DUO) dataset, the A-EIoU model achieved a robust 44.2% Mean Average Precision (mAP@50). Furthermore, under high-noise synthetic stress tests simulating turbid water and human annotation error, the model demonstrated graceful degradation, with mAP dropping minimally to 42.9%. This confirms that A-EIoU learns the organism's true visual features rather than rigid box coordinates, enabling highly reliable, real-time population counting for autonomous underwater vehicles.