Underwater Monocular Vision 3D Reconstruction Based on Cascaded Epipolar

Addressing the challenges of low accuracy in image depth prediction and significant disparity between reconstructed edges and reality in underwater monocular 3D reconstruction, this paper conducts an in-depth analysis of a deep learning network based on Cascaded Epipolar geometry. By employing model pre-training, camera pose inversion, adjusting the format of cascaded convolutions, and modifying the epipolar direction, the model gains the ability to perform 3D reconstruction of underwater monocular visual images. Experiment results on the public dataset Flsea indicates that the Cascaded Epipolar-based method for underwater monocular vision 3D reconstruction offers clearer depth prediction edges and 3D depth information. However, its performance in terms of Mean Squared Error (MSE) shows some deviation comparing to the state-of-the-art UW -GAN.