STAR: Sparse Trained Articulated Human Body Regressor (2008.08535v1)

Abstract: The SMPL body model is widely used for the estimation, synthesis, and analysis of 3D human pose and shape. While popular, we show that SMPL has several limitations and introduce STAR, which is quantitatively and qualitatively superior to SMPL. First, SMPL has a huge number of parameters resulting from its use of global blend shapes. These dense pose-corrective offsets relate every vertex on the mesh to all the joints in the kinematic tree, capturing spurious long-range correlations. To address this, we define per-joint pose correctives and learn the subset of mesh vertices that are influenced by each joint movement. This sparse formulation results in more realistic deformations and significantly reduces the number of model parameters to 20% of SMPL. When trained on the same data as SMPL, STAR generalizes better despite having many fewer parameters. Second, SMPL factors pose-dependent deformations from body shape while, in reality, people with different shapes deform differently. Consequently, we learn shape-dependent pose-corrective blend shapes that depend on both body pose and BMI. Third, we show that the shape space of SMPL is not rich enough to capture the variation in the human population. We address this by training STAR with an additional 10,000 scans of male and female subjects, and show that this results in better model generalization. STAR is compact, generalizes better to new bodies and is a drop-in replacement for SMPL. STAR is publicly available for research purposes at http://star.is.tue.mpg.de.

• The paper presents a novel sparse formulation that localizes pose-driven deformations, reducing parameters to 20% of SMPL’s 4.2 million.
• It decomposes corrective blend shapes into pose and shape factors, using BMI to tailor deformations for individual body characteristics.
• By incorporating over 10,000 additional scans, STAR achieves superior generalization and offers an efficient, SMPL-compatible alternative for 3D human modeling.

An Insight into the STAR Model: Sparse Trained Articulated Human Body Regressor

The paper "STAR: Sparse Trained Articulated Human Body Regressor" introduces a novel model that directly addresses acknowledged shortcomings in the widely-utilized SMPL model for 3D human pose and shape estimation. SMPL, while effective, has notable limitations related to parameter density and the inability to tailor deformations based on both pose and individual body shape characteristics. The STAR model proposes a refined approach that offers improved generalizability and computational efficiency.

Limitations of SMPL and the STAR Model's Innovations

The SMPL model, while seminal in the field, suffers from a large parameter space driven by a dense formulation of pose-corrective offsets. Specifically, every vertex on the mesh in SMPL is influenced by global blend shapes that do not discriminate between geodesically distant body joints. This can lead to spurious long-range mesh deformations and overfitting, despite the use of substantial regularization during training.

In contrast, the STAR model is predicated on the insight that pose deformations are largely local phenomena. To capture this locality, STAR employs a sparse formulation that associates each joint movement only with a concise subset of mesh vertices. This results in a substantial reduction in model parameters to 20% of those for SMPL, specifically lowering from 4.2 million parameters to significantly fewer. The STAR model achieves this by utilizing a novel thresholding function during the learning process, which effectively performs a dense to sparse transition in vertex influence distribution.

Model Formulation and Training

The architecture of STAR relies on a decomposed corrective blend shapes mechanism that separately accounts for pose and shape influences. Thereby it addresses the limitation in SMPL where shape-dependent deformations were inaccurately simplified by being decoupled from the pose. Specifically, STAR introduces shape-dependent deformations informed by the Body Mass Index (BMI), creating blend shapes tailored by both pose and body shape characteristics.

For the integration of differences in individual BMI, STAR nuances its corrective blend shape functions, allowing for more realistic and varied body deformations under different poses. Beyond this, STAR demonstrates a superior capture of the human body's variability through the inclusion of additional 10,000 scans from diverse subjects which surpass the scope of SMPL's original training set (principally the CAESAR dataset from the 1990s).

Numerical and Theoretical Implications

STAR's introduction implies significant benefits for a vast array of applications in computer vision, graphics, apparel, and health sectors. With its 80% reduced parameter burden and improved generalization on unseen data, the model streamlines computational needs while concurrently enhancing fidelity to realistic human anatomy. STAR's focus on spatially local corrective deformation also circumvents the previous challenge in graphics applications where non-local adjustments led to unintentional effects distant from primary joint movements.

Theoretical implications of STAR extend toward pioneering a unified framework that better approximates real-world human body variability, challenging prior notions about joint-wise influence distributions formalized in previous models like SMPL.

Future Developments

The advancements encapsulated in STAR pave the way for subsequent developments that could integrate similar sparse regressor strategies within more comprehensive models such as SMPL-X, which includes the more expressive face and hand features. Additionally, STAR's infrastructure, already modeled to work as a direct SMPL substitute, suggests potential for refining existing applications that rely on SMPL without necessitating fundamental overhauls in the operational pipeline.

STAR offers a notable contribution to computational morphology by redefining parameter efficiency and generalization paradigms within human body modeling. By publicly making this model accessible with an expanded 300-shape principal component space, the authors have positioned STAR as a pivotal tool for continued innovation in 3D human modeling endeavors.