AMASS: Archive of Motion Capture as Surface Shapes

TL;DR

AMASS is a comprehensive human motion dataset that unifies multiple mocap datasets into a standard format using a new conversion method, enabling advanced research and applications in computer vision and animation.

Contribution

The paper introduces AMASS, a large, unified human motion dataset created by converting diverse mocap data into a common surface-based representation using MoSh++, facilitating easier integration and analysis.

Findings

AMASS contains over 40 hours of motion data from 300+ subjects.

MoSh++ effectively converts various mocap datasets into realistic 3D meshes.

The dataset enhances capabilities for animation, visualization, and deep learning training.

Abstract

Large datasets are the cornerstone of recent advances in computer vision using deep learning. In contrast, existing human motion capture (mocap) datasets are small and the motions limited, hampering progress on learning models of human motion. While there are many different datasets available, they each use a different parameterization of the body, making it difficult to integrate them into a single meta dataset. To address this, we introduce AMASS, a large and varied database of human motion that unifies 15 different optical marker-based mocap datasets by representing them within a common framework and parameterization. We achieve this using a new method, MoSh++, that converts mocap data into realistic 3D human meshes represented by a rigged body model; here we use SMPL [doi:10.1145/2816795.2818013], which is widely used and provides a standard skeletal representation as well as a fully rigged surface mesh. The method works for arbitrary marker sets, while recovering soft-tissue dynamics and realistic hand motion. We evaluate MoSh++ and tune its hyperparameters using a new dataset of 4D body scans that are jointly recorded with marker-based mocap. The consistent representation of AMASS makes it readily useful for animation, visualization, and generating training data for deep learning. Our dataset is significantly richer than previous human motion collections, having more than 40 hours of motion data, spanning over 300 subjects, more than 11,000 motions, and will be publicly available to the research community.

Figures

Tables

Table 1. Table 1: Datasets contained in AMASS. We use MoSh++ to map more than 42 42 42 hours of marker data into SMPL parameters, giving a unified format.

	Markers	Subjects	Motions	Minutes
ACCAD [5]	82	20	258	27.22
BioMotion [41]	41	111	3130	541.82
CMU [14]	41	97	2030	559.18
EKUT [29]	46	4	349	30.74
Eyes Japan [6]	37	12	795	385.42
HumanEva [38]	39	3	28	8.48
KIT [29]	50	55	4233	662.04
MPI HDM05 [32]	41	4	219	147.63
MPI Limits [8]	53	3	40	24.14
MPI MoSh [27]	87	20	78	16.65
SFU [2]	53	7	44	15.23
SSM	86	3	30	1.87
TCD Hand [22]	91	1	62	8.05
TotalCapture [42]	53	5	40	43.71
Transitions	53	1	115	15.84
Total	-	346	11451	2,488.01

Equations

$$S(\mathbf{\beta },\mathbf{\theta },\mathbf{\varphi })$$

$$T(\mathbf{\beta },\mathbf{\theta },\mathbf{\varphi })$$

$$\begin{array}{rl}E(\tilde{\mathcal{M}},\mathbf{\beta },{\Theta }_B,{\Theta }_H)& ={\lambda }_D{E}_D(\tilde{\mathcal{M}},\mathbf{\beta },{\Theta }_B,{\Theta }_H)+{\lambda }_{\mathbf{\beta }}{E}_{\mathbf{\beta }}(\mathbf{\beta })\\ & +{{\lambda }_{\mathbf{\theta }}}_B{E}_{{\mathbf{\theta }}_B}({\mathbf{\theta }}_B)+{{\lambda }_{\mathbf{\theta }}}_H{E}_{{\mathbf{\theta }}_H}({\mathbf{\theta }}_H)\\ & +{\lambda }_R{E}_R(\tilde{\mathcal{M}},\mathbf{\beta })+{\lambda }_I{E}_I(\tilde{\mathcal{M}},\mathbf{\beta }).\end{array}$$

$$E_{{\mathbf{\theta }}_H}({\mathbf{\theta }}_H)={{\hat{\mathbf{\theta }}}_H}^T{\Sigma }_{{\mathbf{\theta }}_H}^{-1}{\hat{\mathbf{\theta }}}_H,$$

$$\begin{gathered} {\lambda }_D=600\times b,{\lambda }_{\mathbf{\beta }}=1.25,{{\lambda }_{\mathbf{\theta }}}_B=0.375, \\ {{\lambda }_{\mathbf{\theta }}}_H=0.125,{\lambda }_I=37.5,{\lambda }_R=1e4 \end{gathered}$$

$$\begin{array}{rl}E({\mathbf{\theta }}_B,{\mathbf{\theta }}_H,\mathbf{\varphi })& ={\lambda }_D{E}_D({\mathbf{\theta }}_B,{\mathbf{\theta }}_H,\mathbf{\varphi })\\ & +{{\lambda }_{\mathbf{\theta }}}_B{E}_{{\mathbf{\theta }}_B}({\mathbf{\theta }}_B)+{{\lambda }_{\mathbf{\theta }}}_H{E}_{{\mathbf{\theta }}_H}({\mathbf{\theta }}_H)\\ & +{\lambda }_u{E}_u({\mathbf{\theta }}_B,{\mathbf{\theta }}_H)\\ & +{\lambda }_{\mathbf{\varphi }}{E}_{\mathbf{\varphi }}(\mathbf{\varphi })+{\lambda }_v{E}_v(\mathbf{\varphi }).\end{array}$$

$$E_{\mathbf{\varphi }}(\mathbf{\varphi })={\mathbf{\varphi }}_t^T{\Sigma }_{\mathbf{\varphi }}^{-1}{\mathbf{\varphi }}_t,$$

$$\begin{gathered} {\lambda }_D=400\times b,{{\lambda }_{\mathbf{\theta }}}_B=1.6\times q,{{\lambda }_{\mathbf{\theta }}}_H=1.0\times q, \\ {\lambda }_u=2.5,{\lambda }_{\mathbf{\varphi }}=1.0,{\lambda }_v=6.0, \end{gathered}$$

Full text

1 Introduction

This paper addresses two interrelated goals. First, we develop a method to accurately recover the shape and pose of a person in motion from standard motion capture (mocap) marker data. This enables the second goal, which is to create the largest publicly available database of human motions that can enable machine learning for applications in animation and computer vision. While there have been attempts in both of these directions, existing mocap databases are insufficient in terms of size and complexity to exploit the full power of existing deep learning tools. There are many different mocap datasets available, but pulling them together into a coherent formulation is challenging due to the use of widely varying markersets and laboratory-specific procedures [17]. We achieve this by extending MoSh [27] in several important ways, enabling us to collect a large and varied dataset of human motions in a consistent format (Fig. 1).

MoSh employs a generative model of the body, learned from a large number of 3D body scans, to compute the full 3D body shape and pose from a sparse set of motion capture markers. The results are realistic, but the method has several important limitations, which make it inappropriate for our task. First, MoSh relies on a formulation of the SCAPE body model [13], which is not compatible with existing body representations and graphics software, making it a poor choice for distributing a dataset. We replace SCAPE with the SMPL body model [28], which uses a kinematic tree, has joints, and is based on blend skinning. SMPL comes with a UV map, which allows researchers to generate their own textures for rendering images and video sequences. SMPL is readily available, widely used, and compatible with most game engines and graphics packages. Second, while MoSh captures some soft tissue motions, these are approximate and represented by changing the identity of a subject over time; that is, they are not true soft-tissue deformations. Here we take the dynamic shape space from DMPL [28], which models these soft-tissue deformations from the SMPL model using a shape space learned from 4D body scans of various subjects. We show that we can recover the soft tissue motions realistically from a sparse set of markers. The resulting body shapes and motions look natural and we show that they are metrically accurate. Third, MoSh did not solve for the pose and motion of the hands. Here we add the recent MANO hand model [36], which is compatible with SMPL, and solve for body and hand pose when hand markers are present. This provides richer and more natural animations. Fourth, to fine tune and evaluate our proposed method, we collect a novel dataset, Synchronized Scans and Markers (SSM), that consists of dense 3D meshes in motion, captured with a 4D scanner, together with traditional marker-based mocap. We separate the sequences into training and testing sets, and train the hyperparameters of MoSh++ to minimize the distance between the ground truth 3D scans and the estimated 3D body meshes. We then evaluate the performance of MoSh++ on the test set, demonstrating the accuracy of the method and allowing a quantitative comparison to MoSh.

MoSh++ enables our key goal of creating a large database of human motions. While there are many motion capture datasets available online for research purposes [8, 42, 27, 14, 38, 4, 41, 5, 22, 32], they are often narrowly focused on a particular class of motions. Even the largest ones are too limited in size and variety to support serious deep learning models. Additionally, datasets vary in the format of the data and the kinematic structure of the body, making it hard for researchers to combine them. There have been several efforts to create data supersets [3, 21, 29], but the process of unifying the datasets typically means standardizing to fixed body proportions, which fundamentally alters the data. We argue that a good dataset should capture the articulated structure of the body in a way that is consistent with standard body models so that it can easily be adapted to new problems. Additionally, richness of the original source marker data should be retained as much as possible. It should also be possible to produce high-quality animations that are realistic enough to train computer vision algorithms; that is, the dataset should include full 3D human meshes.

SMPL provides the unifying representation that is independent of the marker set, yet maintains the richness of the original marker data, including the 3D body shape. We know of no other attempt that provides access to full body shape and soft-tissue from mocap data, while also providing accurate body and hand pose. Here we combine $15$ existing motion capture datasets into one large dataset: the Archive of Mocap as Surface Shapes (AMASS). AMASS has $42$ hours of mocap, $346$ subjects, and $11451$ motions. The source datasets all contain varying markersets ranging in size from 37 to 91 markers; AMASS unifies these into a single format. Each frame in AMASS includes the SMPL 3D shape parameters ($16$ dimensions), the DMPL soft tissue coefficients ($8$ dimensions), and the full SMPL pose parameters ($90$ dimensions), including hand articulations. Users who only care about pose can ignore body shape and soft tissue deformations if they wish. Similarly, the SMPL shape space makes it trivial to normalize all bodies to the same shape if users want joint locations normalized to a single shape. Fig. 1 shows a selection of poses and body shapes in the dataset while Fig. 2 illustrates the difference between MoSh++ and traditional mocap. Traditional datasets contain skeletons and/or makers, while the AMASS dataset also provides fully rigged 3D meshes. With MoSh++ it is easy to add more data and we will continue to expand the dataset. We make AMASS available to the research community at http://amass.is.tue.mpg.de, and will support the community in adding new captures as long as they can be similarly shared.

In summary, we provide the largest unified mocap dataset (AMASS) to the community, enabling new applications that require large amounts of training data.

2 Related Work

There is a large literature on estimating skeletal parameters from mocap markers as well as several commercial solutions that solve this problem. As shown by Gorton et al. [17], different solutions use different skeletal models and pre-specified markersets, which makes it hard to unify the existing corpora of marker-based human recordings. Furthermore, all the methods that fit skeletons to data effectively lose rich surface information in the process. We review the most related work: fitting surface models to markers, capturing hands and soft-tissue motion from markers, and previous motion capture datasets.

Surface Models from Markers. To reconstruct bodies from markers, most methods first build a statistical model of body shape [10] or body shape and pose [11, 13, 28]. Allen et al. [10] reconstruct body shape using $74$ landmarks. They do this only for a fixed body pose and the correspondences between the model and the markers are known. The approach cannot deal with arbitrary poses because the model cannot be posed. Anguelov et al. [13] go further by learning a model (SCAPE) of shape and non-rigid pose deformations. Their method requires a dense 3D scan of each subject, which restricts its application to archival mocap datasets of markers.

Loper et al. [27] address some of these limitations with MoSh, and remove the requirement for individual 3D dense scans. MoSh uses a BlendSCAPE body model formulation [19], which is not compatible with standard graphics packages making it sub-optimal for distribution. Furthermore, MoSh does not capture real soft-tissue dynamics, and does not capture hands.

**Hands. ** There is a large body of work on fitting hand models to RGB-D data [39, 40] but here we focus on methods that capture hand motion from sparse markers. Maycock et al. [30] combine an optimal assignment method with model fitting but can capture only hands in isolation from the body and require a calibration pose. Schroder et al. [37] propose an optimization method to find a reduced sparse marker set and, like us, they use a kinematic subspace of hand poses. Alexanderson et al. [9] capture hand motion using sparse markers (3-10). They generate multiple hypotheses per frame and then connect them using the Viterbi algorithm [16]. They can track hands that exit and re-enter the scene and the method runs in real time. However, a new model needs to be trained for every marker set used. Han et al. [18] address the problem of automatically labeling hand markers using a deep network. The above methods, either do not estimate hands and bodies together or do not provide a 3D hand shape.

Soft-tissue motion. Most of the work in the mocap community focuses on minimizing the effect of skin deformations on the marker motions [12, 24]. In some biomechanical studies, the markers have even been fixed to the bones via percutaneous pins [23]. Our work is very different in spirit. We argue that, for animation, such soft-tissue and skin deformation makes captured subjects look alive. In [27] they capture soft-tissue by fitting the parameters of a space of static body shapes to a sparse set of markers. This corresponds to modeling soft-tissue deformation by changing the identity of a person. Instead, inspired by Dyna [35], we use the dynamic shape space of DMPL [28], which is learned from thousands of dynamic scans. This results in more realistic soft-tissue motions with minimal increase in model complexity.

Motion Capture Datasets. There are many motion capture datasets [42, 14, 31, 32, 38, 8, 41, 44] as well as several attempts to aggregate such datasets into larger collections [3, 21, 29]. Previous attempts to merge datasets [21, 29] adopt a common body representation in which the size variation among subjects is normalized. This enables methods that focus on modeling pose and motion in terms of joint locations. On the other hand, such an approach throws away information about how body shape and motion are correlated and can introduce artifacts in retargeting all data to a common skeleton. For example, Holden et al. [21] retarget several datasets to a common skeleton to enable deep learning using joint positions. This retargeting involves an inverse kinematics optimization that fundamentally changes the original data.

Our philosophy is different. We work directly with the markers and not the skeleton, recovering the full 3D surface of the body. There is no loss of generality with this approach as it is possible to derive any desired skeleton representation or generate any desired markerset from the 3D body model. Moreover, having a body model makes it possible to texture and render virtual bodies in different scenes. This is useful for many tasks, including generating synthetic training for computer vision tasks [43].

3 Technical Approach

To create the AMASS dataset, we generalize MoSh in several important ways: 1) we replace BlendSCAPE by SMPL to democratize its use (Sec. 3.1); 2) we capture hands and soft-tissue motions (Sec. 3.2); 3) we finetune the weights of the objective function using cross-validation on a novel dataset, SSM (Sec. 4).

3.1 The Body Model

AMASS is distributed in the form of SMPL body model parameters. SMPL uses a learned rigged template $\mathbf{T}$ with $N=6890$ vertices. The vertex positions of SMPL are adapted according to identity-dependent shape parameters, ${\boldsymbol{\beta}}$, the pose parameters, ${\boldsymbol{\theta}}$, and translation of the root in the world coordinate system, $\gamma$. The skeletal structure of the human body is modeled with a kinematic chain consisting of rigid bone segments linked by joints. Each body joint has 3 rotational Degrees of Freedom (DoF), parametrized with exponential coordinates. We use a variant of SMPL, called SMPL-H [36], which adds hand articulation to the model using a total of $n=52$ joints, where $24$ joints are for the body and the remaining $28$ joints belong to the hands. For simplicity of notation, we include the 3D translation vector $\gamma$ in the pose vector. The pose ${\boldsymbol{\theta}}$ is determined by a pose vector of $3\times{52}+3=159$ parameters. The remaining attributes of the SMPL-H model are the same as SMPL.

We combine SMPL-H with DMPL to obtain a model that captures both hand pose and soft-tissue deformations. For brevity we refer to the combined SMPL-H + DMPL model as SMPL throughout this paper, although this goes beyond any published model.

SMPL modifies the template in an additive way. It applies additive shape, pose, and dynamic blendshapes to a template in a canonical pose and predicts joint locations from the deformed surfaces. The model is

[TABLE]

where $G(\mathbf{T},\mathbf{J},{\boldsymbol{\theta}},\mathbf{W}):\mathbb{R}^{3N}\times{\mathbb{R}}^{|\theta|}\times{\mathbb{R}}^{3K}\times{\mathbb{R}^{4\times{3N}}}\mapsto\mathbb{R}^{3N}$ is a linear blend skinning function that takes vertices in the model in the rest pose $\mathbf{T}$, $K$ joint locations stacked in $\mathbf{J}$, a pose ${\boldsymbol{\theta}}$, and the blend weights $\mathbf{W}$, and returns the posed vertices. The blendshape functions $B_{s}({\boldsymbol{\beta}})$, $B_{p}({\boldsymbol{\theta}})$, and $B_{d}({\boldsymbol{\phi}})$ output vectors of vertex offsets relative to the mean template, $\mathbf{T}_{\mathrm{\mu}}$ (refer to [28], [35] for the detailed explanation of the functions). We call these shape, pose, and dynamic blend shapes respectively. Note that the pose blendshape are a function of the pose ${\boldsymbol{\theta}}$, while ${\boldsymbol{\beta}}$ and ${\boldsymbol{\phi}}$ correspond to linear coefficients that determine the shape and soft-tissue deformation.

SMPL captures the dimensionality of body space more compactly than BlendSCAPE. With only $16$ shape, and $8$ dynamics components, MoSh++ achieves better accuracy than MoSh using $100$ shape components. The number of shape and dynamics coefficients is chosen using the SSM dataset such that MoSh++ does not overfit to mocap markers (see Supplemental Material).

3.2 Model Fitting

Similar to MoSh [27], MoSh++ uses two stages to fit a body model to a sparse markerset. In the following we explain these two stages, while briefly mentioning the overlapping parts for completeness, and highlighting our contributions. For comparability we use a similar notation to the original MoSh paper.

Stage I: Following MoSh, we use a marker parametrization $m(\tilde{\mathbf{m}}_{i},{\boldsymbol{\beta}},{\boldsymbol{\theta}}_{t})$ that maps a latent, pose invariant representation of the markers, $\tilde{\mathbf{m}}_{i}$, to estimate their position in a posed frame, ${\boldsymbol{\theta}}_{t}$. In the first stage, for $F=12$ randomly chosen frames from the subject-specific mocap sequences, given an initial guess for marker-body correspondences, we optimize poses $\Theta={\boldsymbol{\theta}}_{1\ldots F}$, a single shape ${\boldsymbol{\beta}}$, and latent marker positions $\tilde{\mathcal{M}}=\{\tilde{\mathbf{m}}_{i}\}$ to fit the observed marker locations $\mathcal{M}=\{\mathbf{m}_{i,t}\in\mathcal{M}_{t}\}_{1\ldots F}$, where $i$ indexes the markers in a frame; at this stage we exclude soft-tissue deformations. More specifically, similar to MoSh, we optimize the following objective function:

[TABLE]

The data term $E_{D}$ measures distance between simulated markers $m(\tilde{\mathbf{m}}_{i},{\boldsymbol{\beta}},{\boldsymbol{\theta}}_{t})$ and the observed ones $\mathbf{m}_{i,t}$; $E_{{\boldsymbol{\beta}}}$ is a Mahalanobis distance shape prior on the SMPL shape components; $E_{{\boldsymbol{\theta}}_{B}}$ regularizes the body pose parameters; $E_{R}$ encourages the latent markers to remain a prescribed distance $d$ from the body surface (here we use an average value of $d=9.5mm$); and $E_{I}$ penalizes deviations of latent markers from their initialized locations defined by the markerset (see [27] for further details).

In addition to the original terms of MoSh in Eq. 3, we add $E_{{\boldsymbol{\theta}}_{H}}$, which regularizes the hand pose parameters. We project the full hand pose (i.e. $90$ hand parameters) into the $24$-D MANO pose space for both hands and compute the Mahalanobis distance in this space

[TABLE]

where $\hat{{\boldsymbol{\theta}}}$ represents the projection of the pose and $\Sigma_{{\boldsymbol{\theta}}_{H}}$ is the diagonal covariance matrix of the low-D PCA coordinates (see  [36]).

In contrast to MoSh, the $\lambda$ hyper-parameters are determined by line search on the training set of SSM (Sec. 4.2). The data term, $E_{D}$, in Eq. 3 uses a sum of squared distances, which is affected by the number of markers in the observed mocap sequence. This is noteworthy since a standard $46$-markerset was used to determine the $\lambda$ weights during the hyper-parameter search. To deal with varying numbers of markers, we automatically vary the weight of this term by scaling it by a factor $b=46/n$, where $n$ is the number of observed mocap markers.

To help avoid local optima while minimizing Eq. 3, we use Threshold Acceptance method [15] as a fast annealing strategy. Over $4$ annealing stages of graduated optimization, we increase $\lambda_{D}$ by multiplying it by a constant factor $s=2$ while dividing the regularizer weights by the same factor. The weights at the final iteration are as follows:

[TABLE]

The surface distance regularization weight, $\lambda_{R}$, remains constant throughout the optimization. The $24$ hand pose components are added into the optimization only during the final two iterations.

Stage II: In this stage, the latent marker locations and body shape parameters ${\boldsymbol{\beta}}$ of the model are assumed constant over time and the objective at this stage optimizes pose for each frame of mocap in the sequence.

Like MoSh, we add a temporal smoothness term for pose changes, $E_{u}$, to help reduce the effect of jitter noise in the mocap marker data. Yet in contrast to MoSh, we optimize for the soft-tissue deformation coefficients, ${\boldsymbol{\phi}}$. We add a prior and a temporal smoothness terms, $E_{{\boldsymbol{\phi}}}({\boldsymbol{\phi}})$ and $E_{v}({\boldsymbol{\phi}})$ respectively, to regularize the soft-tissue deformations. Then the final objective function for this stage becomes

[TABLE]

The data, body, and hands pose prior terms, $E_{D}$, $E_{{\boldsymbol{\theta}}_{B}}$, and $E_{{\boldsymbol{\theta}}_{H}}$, are the same as described in the first stage. To regularize the soft-tissue coefficients, we add a Mahalonobis distance prior on the $8$ DMPL coefficients.

[TABLE]

where the covariance $\Sigma_{{\boldsymbol{\phi}}}$ is the diagonal covariance matrix from the DYNA dataset [35].

When hand markers are present, MoSh++ optimizes the hand pose parameters in the same way as all the other pose parameters except that we use $24$ dimensions of MANO’s [36] low-dimensional representation of the pose for both hands. In cases where there are no markers present on the hands of the recorded subjects, the hand poses are set to the average pose of the MANO model.

The initialization and fitting for the first frame of a sequence, undergoes a couple of extra steps compared to the rest of the motion. For the first frame, we initialize the model by performing a rigid transformation between the estimated and observed markers to repose the model from its rest pose $\mathbf{T}$ to roughly fit the observed pose. Then we use a graduated optimization for Eq. 6 with only the data and body pose prior terms, while ${\lambda_{{\boldsymbol{\theta}}}}_{B}$ is varied from $[10,5,1]$ times the final weight. Later, for each of the subsequent frames, we initialize with the solution of the previous frame to estimate the pose and soft-tissue parameters.

The per-frame estimates of dynamics and pose after the first frame are carried out in two steps. During the first step, we remove the dynamics and dynamics smoothness terms, and optimize only the pose. This prevents the dynamics components from explaining translation or large pose changes between consecutive frames. Then, we add the dynamics, ${\boldsymbol{\phi}}$, and the dynamics smoothness terms into the optimization for the final optimization of pose and dynamics.

We explain details of tuning the weights $\lambda$ in Sec. 4.2. The velocity constancy weights $\lambda_{u}$ and $\lambda_{v}$ depend on the mocap system calibration and optical tracking quality, data frame rate, and the types of motions. Therefore, these values could not be optimized using just one source of data, so we empirically determined them through experiments on different datasets of varying frame rates and motions. The final weights determined for this stage are:

[TABLE]

Similar to $b$, which adjusts the weight of the data term to varying markersets, $q$ is a weight-balancing factor for the pose prior $\lambda_{{\boldsymbol{\theta}}}$. During a mocap session, markers may get occluded by the body due to pose. If multiple markers of a particular body part are occluded simultaneously, the optimization may result in unreliable and implausible poses, such as the estimated pose shown in Fig. 4 (left). To address this, we introduce a coefficient q=1+\big{(}\frac{x}{|\mathcal{M}|}*2.5\big{)}, where $x$ is the number of missing markers in a given frame, $|\mathcal{M}|$ are the number of total observed markers. This updates the pose prior weight as a factor of the number of missing markers. The more markers that are missing, the higher this weights the pose prior. This term can increase the prior weight by up to a factor of $q=3.5$, in the worse case scenario where $x=|\mathcal{M}|$, and goes down to having no effect, $q=1.0$ when all session markers are visible $x=0$. An example of the pose-estimation with this factor is shown in Fig. 4 (right).

3.3 Optimization and Runtime

Similar to MoSh we use Powell’s gradient based dogleg minimization [34] implemented in the Chumpy [26] auto-differentiation package. Details on the runtime are presented in the Supplementary Material.

4 Evaluation

In order to set the hyperparameters and evaluate the time varying surface reconstruction results of MoSh++, we need reference ground truth 3D data with variations in shape, pose and soft-tissue deformation. To that end, we introduce the SSM dataset (Sec. 4.1) and optimize the weights of the objective functions (Eqs. 3 and 6) using cross-validation on SSM (Sec. 4.2). After optimizing the hyper-parameters, we evaluate the accuracy of MoSh++, e.g. shape reconstruction accuracy (Sec. 4.3), pose, and soft-tissue motion reconstruction (Sec. 4.4) on the test set.

4.1 Synchronized Scans and Markers (SSM)

We use an OptiTrack mocap system [33] to capture subjects with 67 markers; i.e. using the optimized marker-set proposed by MoSh. The system was synchronized to record the mocap data synchronously with a 4D scanning system [1]. See Fig. 5; details are provided in the supplementary material. The dataset consists of three subjects with varying body shapes, performing a total of $30$ different motions. Two of the three subjects were professional models who signed modeling contracts; this allows us to release their 4D scan data, along with the synchronized mocap data for the research community.

We evaluate the accuracy of MoSh++ using the 67 markers, as well as a more standard 46 marker subset of the 67 markers. For both testing and evaluation, we use scan-to-model distances between the 3D scans (our ground truth mesh) of the SSM dataset and the corresponding estimated meshes for each trial of the hyper-parameter search and evaluation. For each reconstructed mocap frame, we take a uniform sampling of 10,000 points of the corresponding synchronized 3D scan and compute the distance from each of these to the closest surface point in our reconstructed mesh. We measure the average of these distances (in $mm$).

4.2 Hyper-parameter Search using SSM

The goal is to set the $\lambda$ weights in Eq. 3 and Eq. 6 to minimize the reconstruction error for the validation data. Grid search complexity grows exponentially with the number of parameters (i.e. $5$ parameters in the case of shape estimation, $4$ in the case of pose estimation). Therefore, we perform line search on each parameter keeping the others fixed.

For the shape estimation stage, the optimization uses $12$ randomly chosen mocap frames from each training subject to estimate shape and marker location for that subject. Instead of choosing a single, unseen pose to evaluate shape accuracy as in [27], we report the average error over the $12$ randomly selected frames from the first stage of Mosh (see Sec. 3.2). Here the duration of the mocap sessions does not matter but variation of body shape among the testing and training subjects is important. Therefore, we separate mocap data of two out of the three SSM subjects as the training set, while keeping the third subject held out for testing and evaluation. We repeat the process $4$ times for the training subjects, using a different random set of $12$ frames for each trial. Validation is performed by running the optimization a fifth time, and initializing with a new randomization seed. We use a line search strategy to determine objective weights [$\lambda_{D}$, $\lambda_{{\boldsymbol{\theta}}}$, $\lambda_{{\boldsymbol{\beta}}}$, $\lambda_{I}$, $\lambda_{R}$] of Eq. 3 by finding a combination of these weights that provide the lowest reconstruction error for the estimated body mesh in the $12$ frames picked during each trial. The final weights are described in Sec. 3.2.

For pose estimation, we separated $20\%$ of the total captured mocap files from the three subjects as a held-out set for testing and evaluation. The first $200$ frames of the rest of the motion files are used for training, leaving the remaining frames (roughly $60\%$ of the training set) for validation. We perform a line search on the objective weights [$\lambda_{D}$, $\lambda_{{\boldsymbol{\theta}}}$, $\lambda_{{\boldsymbol{\phi}}}$] of Eq. 6 and the missing-marker coefficient $q$, obtaining the final weights described in Sec. 3.2.

4.3 Shape Estimation Evaluation

Compared to MoSh, we obtain more accurate results on $SSM$. Fig. 6 (left) shows that the shape estimation accuracy on SSM is $12.1mm$ and $7.4mm$ for MoSh and MoSh++ respectively, when using a standard 46-markerset. Note that we use SSM to determine the optimal number of shape and dynamic coefficients (16 and 8 respectively). Adding more decreases marker error but this overfits to the markers, causing higher error compared with the ground truth shape. Details are in the Supplemental Material.

4.4 Pose and Soft-tissue Estimation Evaluation

We also evaluate the per frame accuracy of pose and soft-tissue motion estimation of MoSh++. Fig. 6 (middle) shows that the pose estimation accuracy on SSM without soft-tissue motion estimation is $10.5mm$ and $8.1mm$ for MoSh and MoSh++ respectively, when using a standard 46-markerset. Similarly, with dynamics terms turned-on, MoSh++ achieves more accurate results than MoSh ($7.3mm$ vs $10.24mm$), Figure 6 (right). The importance of soft-tissue estimation can be observed in Fig. 7. This result is expected since MoSh [27] models soft-tissue motion in the form of changes in the identity shape space of the BlendSCAPE model, whereas MoSh++ fits the proper DMPL space of soft-tissue motions learned from data [28] to mocap markers.

4.5 Hand Articulation

We do not have ground-truth data for evaluating accuracy of hand articulation. Qualitative results of our joint body and hand captures can be seen in Fig. 8. Notice how MoSh++ with hand capture leads to more realistic hand poses. This illustrates that MoSh++ is not limited to the main body but can be extended to capture other parts if a model is available.

5 AMASS Dataset

We amassed in total $15$ mocap datasets, summarized in Table 1. Each dataset was recorded using a different number of markers placed at different locations on the body; even within a dataset, the number of markers varies. The publicly available datasets were downloaded from the internet. We obtained several other datasets privately or recorded them ourselves (Dancers, Transitions, BioMotionLab and SSM). We used MoSh++ to map this large amount of marker data into our common SMPL pose, shape, and soft-tissue parameters. Due to inherent problems with mocap, such as swapped and mislabeled markers, we manually inspected the results and either corrected or held out problems. Fig. 1 shows a few representative examples from different datasets. The supplementary material provides video clips that illustrate the diversity and quality of the dataset. The result is AMASS, the largest public dataset of human shape and pose, including $346$ subjects, $11451$ motions and $42$ hours of recordings and is available to the research community at http://amass.is.tue.mpg.de.

6 Future Work and Conclusions

Future work will extend SSM dataset to include captures with articulated hands. We also intend to extend MoSh++ to work with facial mocap markers. This should be possible using the recently published FLAME model [25], which uses a similar parametrization as the SMPL and MANO models to model facial identity and expressions. Current runtime for MoSh++ is not real-time. However, in principle it should be possible to improve the runtime of MoSh++ significantly by using a parallel implementation of SMPL using frameworks such as TensorFlow [7]. Finally, we see an opportunity to push our approach further to address the problems of missing markers and to exploit the body for fully automatic marker labeling. AMASS itself can be leveraged for this task and used to train models that denoise mocap data (cf. [20]).

In conclusion, we have introduced MoSh++, which extends MoSh and enables us to unify marker-based motion capture recordings, while being more accurate than simple skeletons or the previous BlendSCAPE version. This allowed us to collect the AMASS dataset containing more than $40$ hours of mocap data in a unified format consisting of SMPL pose (with articulated hands), shape and soft-tissue motion. We will continuously incorporate more mocap data contributed by us and the community.

APPENDIX

1 Acknowledgements

We thank Iman Abbasnejad for his immense support and assistance during early development stages of this project, Dr. Senya Polikovsky, Andrea Keller and Haiwen Feng for help in project coordination, motion capture and 4D scanning and data processing, and Talha Zaman, Dr. Javier Romero and Dr. Dimitris Tzionas for their invaluable advice, technical guidance and helpful discussions on body models and optimization techniques.

N. F. Troje held an NSERC Discovery Grant in support of his research and used the Humboldt Research Award granted by the Alexander-von-Humboldt Foundation to fund several visits with his collaborators.

Gerard Pons-Moll has been funded throughout this work by MPI and by the Deutsche Forschungsgemeinschaft (DFG German Research Foundation) project 409792180.

We are grateful to all the authors that have made their mocap datasets available and allowed us to include them into our database:

• •

ACCAD dataset was was obtained from https://accad.osu.edu/research/motion-lab/system-data

• •

The Biomotion Lab dataset (BML) belongs to Dr. Nikolaus F. Troje, who is currently at York University, Toronto, Canada.

• •

The CMU mocap data used in this project was obtained from http://mocap.cs.cmu.edu. That database was created with funding from NSF EIA-0196217.

• •

CMU Kitchen data was obtained from http://kitchen.cs.cmu.edu and the data collection was funded in part by the National Science Foundation under Grant No. EEEC-0540865.

• •

Eyes Japan motion capture data is licensed by mocapdata.com, Eyes, JAPAN Co. Ltd. under the Creative Commons Attribution 2.1 Japan License. To view a copy of this license visit http://creativecommons.org/licenses/by/2.1/jp/.

• •

Human Eva dataset was obtained from http://humaneva.is.tue.mpg.de/. This project was supported in part by gifts from Honda Research Institute and Intel Corporation.

• •

KIT and EKUT data was obtained from KIT Whole Body Motion Database https://motion-database.humanoids.kit.edu/

• •

MPI HDM05 data was was obtained from HDM05 http://resources.mpi-inf.mpg.de/HDM05/

• •

SFU data was was obtained from http://mocap.cs.sfu.ca. The database was created with funding from NUS AcRF R-252-000-429-133 and SFU President’s Research Start-up Grant.

• •

TCD Hands dataset was provided by Trinity College Dublin. The work was sponsored by Science Foundation Ireland as part of the Captavatar, NaturalMovers and Metropolis projects.

• •

Total Capture dataset was obtained from http://cvssp.org/data/totalcapture/. The work was supported by an EPSRC doctoral bursary and InnovateUK via the Total Captureproject, grant agreement 102685. The work was supported in part by the Visual Media project(EU H2020 grant 687800) and through donation of GPU hardware by Nvidia corporation.

• •

The Transitions dataset was provided by Dr. Mazen Al Borno, Dr. Javier Romero and Andrea Keller who contributed to the design and capture of this dataset.

Conflict of Interest Disclosure: N. Mahmood is a founder and shareholder of Meshcapade GmbH, which is commercializing body shape technology. N. Mahmood is with Meshcapade GmbH but this work was performed primarily at the MPI for Intelligent Systems. Michael J. Black has received research gift funds from Intel, Nvidia, Adobe, Facebook, and Amazon. While MJB is a part-time employee of Amazon, his research was performed solely at, and funded solely by, MPI. MJB is also an investor in Meshcapde GmbH.

2 Optimization and Runtime

Section 3.3 of the main paper briefly describes the implementation of MoSh++. All methods were tested on CPU using an early 2015 edition MacBook Pro with Intel Core i7, running at 3.1GHz and 16GB RAM. Runtimes for each step depend heavily on the frame rate and pose variation of the motion. For sequences in the SSM dataset, average runtimes of MoSh++ are as follows:

• •

Stage I: about $25$ min/motion sequence

• •

Stage II without dynamics: about $0.5$ sec/frame

• •

Stage II with dynamics: about $2$ sec/frame

3 Data Collection

Section.4.1 of the main paper presents the SSM (Synchronized Scans and Markers) dataset. To record this dataset we use an optical motion capture system synchronized and calibrated together with a high resolution 4D scanning system.

We used an OptiTrack motion capture system (NaturalPoint, Inc. DBA OptiTrack. Corvallis, OR) [33] consisting of 24 Optitrack Prime 17W optical mocap cameras. Each subject was fitted with 67 reflective mocap markers based on the optimized marker-set layout proposed in [27]. The subjects wore minimal clothing to avoid artefacts due to sliding of cloth. The markers were placed directly on the skin of the subjects wherever possible.

The motion capture system was synchronized to be triggered with a 3dMD 4D body scanning system (3dMD LLC, Atlanta, GA) [1]. The 4D scanner is capable of capturing high-resolution 3D scans of a person at 60 frames per second. The 4D system uses 22 pairs of stereo cameras, 22 color cameras and 34 speckle projectors and arrays of white-light LED panels.

4 Model Size

Section 3.1 of the main paper describes the SMPL body model incorporated in the MoSh++ pipeline. We experimented with varying numbers of SMPL shape components and DMPL dynamic components to find the optimal number to use to capture shape and soft-tissue motion. We found that $|{\boldsymbol{\beta}}|=16$, and $|{\boldsymbol{\phi}}|=8$ do the best job of minimizing error on the held-out validation set and also produce natural looking soft-tissue deformations. Given a limited number of mocap markers, allowing more shape variation results in overfitting (see Fig. 9).

5 Diversity and Quality

We provide a video to illustrate the variations in the motions in AMASS and the quality of reconstructed body surface deformations from the mocap markers. Please see the video. Note that AMASS captures a wide range of body shapes and motions.