TL;DR
This paper introduces a user-friendly system for generating realistic 3D human body models from a minimal set of anthropometric measurements. By employing a novel feature-selection-based local mapping technique and an imputation method for missing data, the system achieves accurate body reshaping with significantly less user effort and without requiring expensive scanning equipment.
What Problem Does This Paper Address?
Creating accurate and realistic 3D human body models is crucial for applications like virtual try-on, person identification, and health monitoring. However, existing methods often rely on complex and costly measurement devices like high-end scanners or range cameras, making them inaccessible for everyday users. Furthermore, these methods can be cumbersome, requiring detailed manual input or segmentation of body parts. The challenge is to develop a system that can generate high-quality 3D body shapes from easily obtainable, limited anthropometric data, making it practical for widespread use.
Methods
The proposed system comprises three main modules: an Imputer, a Selector, and a Mapper, operating in both offline and online stages.
- Imputer: This module handles missing user input data. It leverages the Multivariate Imputation by Chained Equations (MICE) technique to fill in incomplete anthropometric measurements, ensuring a complete set of parameters for the reshaping process.
- Selector: This is the core of the novelty. It employs a "feature-selection-based local mapping technique" to automatically identify the most relevant anthropometric parameters for each specific facet (triangle) of the 3D body mesh. This is achieved by learning "relevance masks" and "mapping matrices" offline. The technique analyzes the statistical relationship between parameters and mesh deformations, recursively eliminating less important features. This avoids manual parameter binding to body parts, a limitation of prior work like the SPRING model.
- Mapper: In the online stage, this module takes the imputed anthropometric parameters and uses the learned relevance masks and mapping matrices from the Selector to generate the final 3D body mesh. It maps the selected parameters to deformation matrices for each facet, which are then used to reconstruct the vertex positions of the mesh.
Datasets
The paper utilizes a dataset containing 1531 female and 1517 male meshes, each with a resolution of 12500 vertices and 25000 facets. This dataset is described as pose-independent. The authors also mention using the CAESAR dataset [11] as a source of opportunities for representing human bodies with 3D models.
Key Insights and Ideas
- Minimal Input, Maximum Output: The system's ability to generate accurate 3D bodies from just 3-5 measurements is a significant step towards user-friendliness and accessibility.
- Automatic Feature Selection: The "feature-selection-based local mapping" is a key innovation. It automates the process of identifying which body measurements are most important for shaping specific parts of the body, eliminating the need for manual feature engineering and potentially leading to more robust models.
- Local Mapping for Robustness: By analyzing deformations on a facet-by-facet basis, the system can better capture local shape variations and avoid the global distortions seen in some other methods when specific parameters are changed.
- Imputation for Practicality: Using MICE to handle missing data makes the system more practical for real-world scenarios where users might not have all measurements readily available.
Key Results
The paper reports superior performance compared to state-of-the-art methods like global mapping and the SPRING model.
- Anthropomorphic Evaluation: The proposed system achieved lower Mean Absolute Errors (MAE) for reconstructing anthropometric parameters across both male and female datasets, indicating higher accuracy. For instance, the "gluteal hip" parameter showed MAEs of 9.5/10.4 (female/male) for the proposed system, compared to 20.6/22.3 for global mapping and 10.3/13.3 for SPRING.
- Qualitative Results: Visual examples demonstrate that when a specific parameter (like gluteal hip or chest circumference) is increased, the proposed system and the SPRING model adjust the relevant body parts while keeping other parts relatively unchanged. In contrast, the global mapping method sometimes caused unintended changes in unrelated body regions.
- User Study: A user study with 68 volunteers indicated that the proposed system was preferred for generating resembling body meshes, especially when considering different body types (Mesomorph, Ectomorph, Endomorph).
Limitations and Future Work
- Limitations: The paper acknowledges that while the local mapping approach is effective, combining it with global methods could lead to even faster and more accurate systems. The current system focuses on local mapping for shape modeling.
- Future Work: The authors plan to explore combining local and global methods to achieve a better balance of speed and accuracy in future work.
Summary updated