VolHuMe: a High-Resolution Large Scale Dataset of Volumetric Human Meshes

Technical Summary of VolHuMe: A High-Resolution Large Scale Dataset of Volumetric Human Meshes

Problem Statement
Current datasets for 4D human reconstruction face a fundamental trade-off between scale, capture quality, and annotation richness. "In-the-wild" datasets often suffer from noisy imagery, sparse viewpoints, and unconstrained lighting, while synthetic datasets lack fine-grained geometric fidelity. Conversely, controlled environment datasets that offer high-quality geometry often cover only a small number of subjects or rely on distant, full-body captures. This reliance on distant views limits the reconstruction of fine-grained, close-up details such as facial expressions, hand articulation, and clothing textures. Furthermore, existing resources frequently lack the comprehensive combination of modalities required for robust training: high-resolution textured meshes, dense point clouds, multi-view RGB/depth data, and accurate parametric model fittings (e.g., SMPL-X) simultaneously.

Methodology and Data Acquisition
To address these gaps, the authors introduce VolHuMe, a large-scale dataset comprising high-quality 4D human scans. The data was acquired using the Mantis Vision Volumetric Capture System, a state-of-the-art studio setup.

• Hardware: The system features a cylindrical capture volume (2.5m diameter, 3m height) equipped with 32 modules. Each module contains two coplanar RGB cameras and one monochromatic camera for structured light depth estimation, totaling 64 RGB cameras (2456 × 2054 pixels) and 32 depth cameras (1280 × 1024 pixels).
• Capture Strategy: Unlike traditional full-body setups, VolHuMe utilizes a close-range, partial-view capture strategy. Cameras are positioned 80–120 cm from the subject. This proximity allows for the highest possible level of detail after merging views, prioritizing fine-grained body-part resolution over distant full-body coverage.
• Subjects and Duration: The dataset includes 104 diverse adult subjects (40% female, 60% male, ages 18–70). Each subject was recorded for one minute, generating a total of 156,000 annotated frames.
• Annotations: The dataset provides extensive ground truth, including:
  • High-resolution meshes and rigged meshes ready for animation.
  • Dense point clouds.
  • Multi-view RGB and depth images.
  • Garment segmentation.
  • SMPL-X fittings: Validated with a skin error of 4.73 mm and a penetrating clothing error of 20.05%, meeting industry standards for live human measurement.
  • FLAME model fittings: Applied to facial features for enhanced accuracy.

Key Contributions

1. High-Resolution Close-Range Dataset: VolHuMe is the first large-scale dataset to prioritize close-range, high-resolution captures of body parts, preserving fine-grained geometric and textural details often lost in distant full-body scans.
2. Comprehensive Ground Truth: It offers a complete set of annotations (meshes, point clouds, SMPL-X/FLAME fittings, and multi-view imagery) that is rarely available in a single resource, enabling diverse research tasks.
3. Sparse Multi-View Benchmark: The dataset challenges current state-of-the-art (SOTA) methods by providing a sparse multi-view setup (32 modules) that demands high-fidelity reconstruction from partial observations, rather than relying on dense camera arrays.

Experimental Results
The authors benchmarked VolHuMe on 3D and 4D reconstruction tasks to evaluate SOTA methods and expose current limitations.

• View Synthesis and Mesh Estimation:

  • Standard NeRF and 3D Gaussian Splatting (3DGS) methods failed to converge on raw data due to viewpoint sparsity and uniform backgrounds. Custom variants involving background compositing were required for stable training.
  • Photogrammetry-based methods achieved the best mesh reconstruction accuracy but still fell short of the ground-truth mesh detail.
  • Radiant Foam achieved the best view-synthesis performance (PSNR 39.20, SSIM 0.981), outperforming both NeRF and 3DGS approaches.
  • 3DGS methods showed improved recovery of mid-frequency textures compared to NeRF but tended to over-smooth fine details.

• 4D Human Reconstruction:

  • GPS-Gaussian: When evaluated on VolHuMe, this method achieved a PSNR of 36.85, significantly outperforming its results on the THuman2.0 dataset (25.57). This indicates that the higher-fidelity ground truth in VolHuMe leads to better downstream reconstruction.
  • Animatable Gaussians: Achieved a PSNR of 29.59 on VolHuMe, comparable to results on the dense-capture ActorHQ dataset (30.36), despite VolHuMe using only 32 sparse modules versus ActorHQ's 160 cameras.
  • Spatial Sampling Analysis: A controlled experiment in Blender demonstrated that the close-range partial views of VolHuMe yield 1,582,839 pixels of effective spatial coverage, compared to 926,774 pixels from a distant 4K full-body image. This confirms that close-range setups provide superior spatial sampling even with fewer cameras.

Significance and Claims
The paper claims that VolHuMe serves as a challenging benchmark that exposes the limitations of current evaluation testbeds. The authors assert that:

1. Close-range partial observations can achieve reconstruction quality on par with, or superior to, dense full-body capture setups in terms of effective spatial sampling and detail recovery.
2. Existing SOTA methods struggle to fully exploit fine-grained geometric and textural details when faced with sparse multi-view inputs, often requiring auxiliary preprocessing (like background compositing) to function.
3. The dataset provides a necessary resource for advancing high-fidelity human reconstruction, rendering, and view synthesis, particularly for tasks requiring detailed articulation of hands, faces, and clothing.

The authors maintain a modest stance, noting that while their dataset enables new capabilities, current methods still require significant adaptation to handle the specific challenges of sparse, close-range data without losing detail. They position VolHuMe not as a solved problem, but as a rigorous testbed to drive future research in 4D human modeling.