MuS-Polar3D: A Benchmark Dataset for Computational Polarimetric 3D Imaging under Multi-Scattering Conditions

Puyun Wang, Kaimin Yu, Huayang He, Xianyu WuarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces MuS-Polar3D, a novel benchmark dataset for underwater 3D imaging under challenging multi-scattering conditions. By providing diverse turbidity levels, multiple viewpoints, and high-precision ground truth, it enables more robust and fair evaluation of polarization-based 3D reconstruction methods, addressing a critical gap in existing benchmarks.

What Problem Does This Paper Address?

Underwater 3D imaging is crucial for many applications, but turbid water, filled with suspended particles, severely degrades image quality through scattering. Polarization-based imaging offers a promising solution by exploiting how light polarization changes when scattered, allowing for better background suppression and distinct advantages in murky environments. However, existing public datasets for this task lack diversity in scattering conditions and viewpoints, making it difficult to compare different methods fairly and hindering the development of truly robust algorithms that can generalize across various underwater scenarios. This limits progress in areas like underwater navigation, robotics, and scientific exploration.

Methods

The authors propose a two-stage framework for underwater 3D reconstruction under scattering: first, descattering to remove the effects of turbidity, and second, 3D reconstruction using the enhanced images. They leverage polarization information in both stages. For dataset construction, they meticulously controlled the scattering conditions by adding varying amounts of emulsion to water. They also developed a pipeline for acquiring multi-view polarization images, obtaining high-precision 3D mesh models using a structured-light scanner, and generating ground-truth normal maps and masks. To analyze the dataset, they quantified object size and texture richness using metrics like effective pixel ratio and Local Binary Patterns (LBP), and then clustered objects based on these properties and material types.

Datasets

The core contribution is the introduction of the MuS-Polar3D dataset. This benchmark dataset comprises:

  • 42 distinct objects with diverse materials (resin sculptures, glazed ceramics, unglazed ceramics) and textures.
  • Seven quantitatively controlled scattering conditions (from clear to highly turbid).
  • Up to five viewpoints for each object.
  • High-precision 3D mesh models (±0.05 mm accuracy).
  • Ground-truth normal maps and foreground masks.
  • 847 processed samples in total.

This dataset is designed to support various vision tasks, including descattering, normal estimation, object segmentation, and 3D reconstruction, and aims to bridge the gap between single-view shape-from-polarization (SfP) methods and multi-view 3D reconstruction paradigms.

Key Insights and Ideas

  • Polarization is key for both descattering and 3D reconstruction: The paper emphasizes that polarization cues are not only useful for suppressing scattering but also for inferring structural information, making it a powerful tool for the entire underwater 3D imaging pipeline.
  • Controlled scattering is crucial for robust evaluation: The ability to precisely control turbidity levels in the dataset allows for a systematic analysis of how scattering affects reconstruction accuracy and how well different algorithms generalize.
  • Multi-view polarization is essential for modern 3D reconstruction: By providing multi-view data, MuS-Polar3D enables fair comparisons between traditional polarization-based methods and state-of-the-art multi-view techniques like SDF-based approaches.
  • Texture complexity significantly impacts reconstruction errors: The analysis reveals a strong correlation between high-frequency texture regions and reconstruction errors, suggesting that while descattering helps, intrinsic object geometry and texture remain critical factors.

Key Results

  • Improved 3D Reconstruction Accuracy: Experiments show that incorporating descattered polarization images as input to 3D reconstruction networks consistently improves accuracy across various baseline methods.
  • AttentionU2Net Performance: The AttentionU2Net method, when combined with descattering, achieved the best overall performance, demonstrating the lowest Mean Angular Error (MAE) in 3D reconstruction.
  • Robustness to Scattering: The proposed two-stage approach, particularly with methods like DehazeFormer for descattering, shows significant robustness, with performance degradation being more gradual compared to raw images as scattering increases.
  • Quantitative Metrics: The paper reports quantitative evaluations using metrics like PSNR, SSIM, LPIPS, NIQE for descattering, and MAE for 3D reconstruction, providing concrete evidence of the effectiveness of their methods and dataset. For instance, the best mean angular error achieved was 15.49°.

Limitations and Future Work

  • Limitations: While the dataset is comprehensive, the authors acknowledge that the current evaluation of 3D reconstruction primarily focuses on normal estimation. The impact of descattering on other downstream tasks like object detection or segmentation is not extensively explored within this paper.
  • Future Work: The authors suggest future research directions including:
  • Further enhancing high-frequency geometric details in reconstructions.
  • Exploring deeper integration of polarization cues with multi-view geometric information.
  • Extending the dataset's utility to other vision tasks like novel view synthesis, material property estimation, and salient object detection.
  • Investigating the potential for single-view normal estimation methods to be compared with multi-view approaches under a unified framework.

Summary updated

BibTeX
@misc{wang2025mus,
  author = {Puyun Wang and Kaimin Yu and Huayang He and Xianyu Wu},
  title = {{MuS-Polar3D: A Benchmark Dataset for Computational Polarimetric 3D Imaging under Multi-Scattering Conditions}},
  year = {2025},
  eprint = {2512.21513},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.21513},
}
Figure