Underwater Surface Normal Reconstruction via Cross-Grained Photometric Stereo Transformer

Abstract

Modern ocean research necessitates high-precision 3-D underwater data acquisition. Photometric stereo is a critical technique for recovering high-resolution, dense surface normals of textureless objects, such as the seabed and underwater pipelines. This technique is fundamental for underwater robots engaged in ocean exploration and operational tasks. Traditional underwater photometric stereo methods account for distributed underwater media, such as light scattering. However, the deployment of devices in complex underwater environments (e.g., ocean currents) often results in misalignment and jitter among photometric stereo images. These challenges lead to inaccuracies in matching-based methods, particularly due to the lack of texture and varying illumination conditions. To address these issues, we propose the Cross-Grained Transformer Photometric Stereo (CGT-PS) Network. CGT-PS is designed to directly manage misaligned pixels caused by underwater jitter in an end-to-end manner. The proposed method consists of two main components: the local-grained and global-grained modules. The local-grained module utilizes a Shift operation to adjust pixels within a single-pixel span, effectively mitigating misalignment caused by motion without increasing computational cost. In contrast, the global-grained module performs nonlocal fusion learning, leveraging distant features to enhance the extraction of intricate structural details, cast shadows, and interreflection regions. Ablation studies confirm the efficacy of the proposed modules. Extensive experiments on photometric stereo benchmark data sets and real underwater photometric stereo samples demonstrate that our method achieves superior performance.