Abstract:In deep-sea environments, acquiring the three-dimensional (3D) morphology of seafloor polymetallic nodules faces numerous challenges: On one hand, constrained by the high costs of deep-sea diving operations and imaging conditions, typically only top-down (overhead) images can be obtained, resulting in severely limited observational information. On the other hand, the scarcity of real 3D scanned data makes it difficult to support the training of deep learning models. To address these issues, this paper proposes a single-view-based method for 3D reconstruction and volume estimation of deep-sea polymetallic nodules. First, to tackle the problem of insufficient training data, a 3D dataset named PMN800 containing 800 synthetic samples was constructed using the 3D modeling software Blender, covering various morphologies and texture characteristics of polymetallic nodules. Subsequently, the SyncDreamer diffusion model is employed to generate multi-view consistent images from single-view inputs. The neural reconstruction framework NeuS is then utilized to fuse these multi-view images into a high-fidelity 3D mesh model. Finally, volume estimation is achieved using a triangulationbased approach. Synthetic data are leveraged for auxiliary training, and the method is tested on real polymetallic nodules. In multi-view consistency evaluation, the proposed method demonstrates a 26.7% improvement in peak signal-to-noise ratio (PSNR) compared to the original model. In subjective quality assessment, fidelity and consistency scores increase by 0.95 and 1.20, respectively, with overall performance surpassing existing mainstream algorithms. Experimental results for volume measurement reveal average relative errors of 7.65% on the synthetic dataset and 9.41% on real deep-sea polymetallic nodule data, validating the method’s effectiveness and reliability in volume measurement of polymetallic nodules.