HuSc3D: Human Sculpture dataset for 3D object reconstruction

Weronika Smolak-Dyzewska, Dawid Malarz, Grzegorz Wilczynski, Rafal Tobiasz, Joanna Waczynska, Piotr Borycki, Przemyslaw Spurek

Introduction

Recent advancements in 3D reconstruction, particularly Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have demonstrated remarkable results on standard benchmarks. However, these benchmarks often rely on synthetic data or professionally curated captures that do not reflect the challenges of real-world, "casual" photography. The HuSc3D dataset is introduced to bridge this gap, providing a rigorous stress test for 3D reconstruction models using highly detailed, monochromatic sculptures captured under realistic "in-the-wild" conditions.

The HuSc3D Dataset

The core of this work is a collection of six scenes featuring sculptures by artist Paweł Althamer. These objects were selected specifically for their challenging geometric properties: fine perforations, low-texture white surfaces, and complex silhouettes that are difficult for standard algorithms to distinguish from backgrounds.

The six sculptures of the HuSc3D dataset: Paulina, Wiktor, Hanna, Tosia, Jarosław, and Kacper.

The dataset is categorized by the specific technical hurdles each scene presents:

  • Sparse Viewpoints: The Kacper and Paulina scenes contain only 66 and 104 photos respectively, testing a model's ability to generalize from limited data.
  • High-Resolution Processing: Jarosław and Hanna feature imagery up to 8K resolution, highlighting computational bottlenecks in current reconstruction pipelines.
  • Photometric Inconsistency: Tosia and Wiktor exhibit significant Automatic White Balance (AWB) shifts, resulting in inconsistent color casts across frames.

Benchmarking Methodology

The authors evaluated six state-of-the-art reconstruction methods to establish a baseline for the dataset: 2D Gaussian Splatting (2DGS), Mip-Splatting, Nerfacto (a NeRF implementation), Splatfacto (3DGS), Splatfacto-MCMC, and Instant-NGP. Evaluation metrics included standard signal quality (PSNR), structural similarity (SSIM), and perceptual similarity (LPIPS), alongside practical considerations like training time and rendering speed (FPS).

Key Findings

The Impact of White Balance

One of the most significant findings is the sensitivity of implicit representations to lighting changes. NeRF-based methods like Nerfacto tend to produce blurry results when faced with AWB shifts, as they attempt to average out the color cast. In contrast, Gaussian Splatting methods can reconstruct the geometry but often introduce artifacts as they struggle with the sudden chromatic changes.

Comparison of white balance shifts in the dataset, showing the reddish tint that challenges reconstruction algorithms.

Textureless Geometry and Sparse Data

The monochromatic nature of the sculptures poses a "textureless trap." Without distinct visual anchors, models struggle to find consistent features. This is particularly evident in sparse-view scenes where NeRF-based methods frequently failed to converge, whereas explicit, point-based representations (GS) proved more robust in maintaining structural integrity.

Performance Leaders

Splatfacto-MCMC emerged as the top performer in terms of qualitative metrics, achieving an average PSNR of 32.45. However, this accuracy comes at a cost; it was the slowest to train and offered lower rendering speeds (~82 FPS) compared to standard 3DGS (~165 FPS). Meanwhile, grid-based methods like Instant-NGP faced scalability issues, encountering Out-of-Memory (OOM) errors on high-resolution scenes.

Conclusion

HuSc3D exposes the limitations of current 3D reconstruction models when moved out of the lab and into realistic capture scenarios. While explicit representations like Gaussian Splatting show promise in handling sparse and low-texture data, significant work remains in addressing photometric inconsistencies and the computational demands of high-resolution imagery. Future expansions of the dataset will aim to combine these challenges to further push the boundaries of robust 3D reconstruction.

Summary updated

Abstract

3D scene reconstruction from 2D images is one of the most important tasks in computer graphics. Unfortunately, existing datasets and benchmarks concentrate on idealized synthetic or meticulously captured realistic data. Such benchmarks fail to convey the inherent complexities encountered in newly acquired real-world scenes. In such scenes especially those acquired outside, the background is often dynamic, and by popular usage of cell phone cameras, there might be discrepancies in, e.g., white balance. To address this gap, we present HuSc3D, a novel dataset specifically designed for rigorous benchmarking of 3D reconstruction models under realistic acquisition challenges. Our dataset uniquely features six highly detailed, fully white sculptures characterized by intricate perforations and minimal textural and color variation. Furthermore, the number of images per scene varies significantly, introducing the additional challenge of limited training data for some instances alongside scenes with a standard number of views. By evaluating popular 3D reconstruction methods on this diverse dataset, we demonstrate the distinctiveness of HuSc3D in effectively differentiating model performance, particularly highlighting the sensitivity of methods to fine geometric details, color ambiguity, and varying data availability--limitations often masked by more conventional datasets.
BibTeX
@inproceedings{smolakdyzewska2026husc,
  author = {Weronika Smolak-Dyzewska and Dawid Malarz and Grzegorz Wilczynski and Rafal Tobiasz and Joanna Waczynska and Piotr Borycki and Przemyslaw Spurek},
  title = {{HuSc3D: Human Sculpture dataset for 3D object reconstruction}},
  booktitle = {ICCS},
  year = {2026},
  doi = {10.1007/978-3-032-29924-6\_54},
}
Figure