Real-Time and Resource-Efficient Multi-Scale Adaptive Robotics Vision for Underwater Object Detection and Domain Generalization

Underwater robotic vision encounters numerous challenges posed by complex environments and varying lighting conditions. Meeting these challenges requires solutions that not only deliver accuracy but also demonstrate adaptability. In this paper, we present MARS (Multi-Scale Adaptive Robotics Vision), a pioneering approach to underwater object detection that prioritizes real-time performance and resource efficiency-essential attributes for underwater robotic systems. Leveraging a well-established object detection architecture, MARS integrates Domain-Adaptive Multi-Scale Attention (DAMSA), enhancing both detection accuracy and adaptability to diverse underwater domains. During training, DAMSA employs domain class-based attention, allowing the model to learn and prioritize features specific to different underwater environments. Extensive evaluation across diverse underwater datasets underscores the effectiveness of MARS. On the original dataset, MARS achieves an impressive mean Average Precision (mAP) of 58.57%, demonstrating its proficiency in detecting critical underwater objects such as echinus, starfish, holothurian, scallop, and waterweeds. This capability positions MARS as a promising solution for applications in marine robotics, marine biology research, and environmental monitoring. Moreover, MARS exhibits exceptional resilience to domain shifts. When evaluated on an augmented dataset incorporating various enhancements, MARS delivers a commendable mAP of 36.16%, showcasing its robustness and adaptability in recognizing objects across diverse underwater conditions. The source code for MARS is publicly available on GitHub at https://github.com/LyesSaadSaoud/MARS-Object-Detection/.