Towards Lightweight Underwater Depth Estimation

Underwater depth estimation is crucial in the applications of marine robotics. It can provide environment information for target tracking, robot navigation, and 3D reconstruction of underwater terrain. Existing works transform underwater images into in-air conditions to adapt methods that are designed for natural images. However, this may result in expensive computational resources. To overcome this limitation, we propose a lightweight knowledge distillation framework for underwater depth estimation. We utilize a powerful model designed for underwater images as the teacher model and a lightweight CNN model as the student model. We distill global features to enable the student to acquire both local and global information, thereby improving estimation performance. Our framework includes a global transformation module for efficient global feature distillation and a global-local fusion module to combine local and global information for final estimation. Experimental results on the FLSea dataset demonstrate that our student model is lighter than the teacher model while outperforming lightweight in-air models. Our network is 60% lighter than the teacher model and achieves a 3.1% improvement in the δ
1
metric compared to the lightweight in-air model.