Generative AI for Human 3D Body Emotions: A Dataset and Baseline

Methods

Ciprian Paduraru

, Petru-Liviu Bouruc

and Alin Stefanescu

1,2

Department of Computer Science, University of Bucharest, Academiei Street 14, Bucharest, Romania

Institute for Logic and Data Science, Romania

{ciprian.paduraru, petru-liviu.bouruc, alin}@unibuc.ro

Keywords:

Generative AI, Body Emotions, Parametric Models, Animations.

Abstract:

Accurate and expressive representation of human emotions in 3D models remains a major challenge in various

industries, including gaming, ﬁlm, healthcare, virtual reality and robotics. This work aims to address this

challenge by utilizing a new dataset and a set of baseline methods within an open-source framework developed

to improve realism and emotional expressiveness in human 3D representations. At the center of this work is the

use of a novel and diverse dataset consisting of short video clips showing people mimicking speciﬁc emotions:

anger, happiness, surprise, disgust, sadness, and fear. The dataset was further processed using state-of-the-

art parametric body models that accurately reproduce these emotions. The resulting 3D meshes were then

integrated into a generative pose generation model capable of producing similar emotions.

1 INTRODUCTION

This work addresses the challenge of enhancing the

emotional expressiveness of 3D human models by

proposing a novel framework. The goal is to gen-

erate 3D representations of human bodies that accu-

rately mimic the emotional expression of real life. To

achieve this, we ﬁrst needed to identify a dataset that

is expressive enough to allow the extraction of 3D

meshes in different poses corresponding to different

emotions.

Dataset. While there are several datasets available

online, they either do not capture the whole body, as

in datasets such as (Sun et al., 2021), (Mollahosseini

et al., 2017) and (Zadeh et al., 2018a), or they focus

on actions rather than emotions, such as clapping in

the (Zadeh et al., 2018b) dataset.

• We have developed a unique dataset that captures

the full body posture of people expressing six spe-

ciﬁc emotions in real time, including the transi-

tions between a normal (relaxed) posture and each

of these postures. This dataset overcomes the lim-

itations of existing datasets by capturing the full

expression cycle and allowing for natural emo-

tional expressions.

• Adapted several of the state-of-the-art methods

that process poses and meshes in both 2D and 3D

space and assembled them into a pipeline process.

• Proposed a new generative process that starts from

any body mesh (even outside the dataset) of a per-

son and generates one of the six emotions our

models are trained on: anger, happiness, surprise,

disgust, sadness and fear.

• The methods, dataset, and experiments are pub-

lished as open source on https://github.com/

unibuc-cs/3DHumanEmotionsGenerator for fur-

ther research in academia and industry.

2 RELATED WORK

The representation of human emotions in the digital

realm has advanced signiﬁcantly in recent decades.

Early efforts, such as those in (Vrajitoru, 2006),

explored digital chatterbots and NPCs capable of

human-like interactions. With the advent of technolo-

gies like deep learning and generative AI, research

has expanded, as shown in (Li et al., 2023) and (Park

et al., 2023), which trained AI models inspired by

neuroscience and simulated interactive bot environ-

ments. Current research emphasizes large language

models (LLMs), which excel at complex tasks and

conversational agent simulations.

Efforts to replicate human 3D representations

have also evolved. Early works (Pons-Moll et al.,

2015), (Kolotouros et al., 2019), (Anguelov et al.,

646

Paduraru, C., Bouruc, P.-L. and Stefanescu, A.

Generative AI for Human 3D Body Emotions: A Dataset and Baseline Methods.

DOI: 10.5220/0013168700003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 646-653

ISBN: 978-989-758-737-5; ISSN: 2184-433X

2005) laid the foundation, with SCAPE (Anguelov

et al., 2005) offering a detailed model for represent-

ing human bodies in 3D. SCAPE combines shape-

and posture-dependent deformations, utilizing prin-

cipal component analysis (PCA) for efﬁciency. Ex-

tensions like Blend-SCAPE (Hirshberg et al., 2012)

introduced smoother transitions, paving the way for

models like SMPL (Loper et al., 2015) and SMPLify

(Bogo et al., 2016), which simplify and improve 3D

human modeling.

Alternative approaches, such as diffusion-based

models, emerged in parallel. ScoreHMR (Stathopou-

los et al., 2024) leverages diffusion models and score-

based learning to overcome optimization challenges,

enhancing 3D mesh recovery. It reﬁnes noisy esti-

mates of 3D meshes through iterative denoising, in-

corporating multi-view reﬁnement and motion data

for dynamic scenes, achieving high accuracy.

Pose estimation, a critical challenge in 3D hu-

man modeling, is addressed by OpenPose (Cao et al.,

2019), (Zhao et al., 2024). Its two-stage architecture

combines CNN-generated conﬁdence maps and part

afﬁnity ﬁelds (PAFs) to connect keypoints into skele-

tons, supporting multi-person scenarios and hand and

facial point recognition.

The novelty of this research lies in integrating and

adapting these techniques to develop a comprehensive

system for realistic whole-body human emotion rep-

resentation using a new dataset, addressing gaps not

tackled by existing models.

3 THEORETICAL

FOUNDATIONS

3.1 Representation of the Human Body

The SMPL (Loper et al., 2015) model represents the

human body as a 3D mesh consisting of about 6,890

vertices, which corresponds to about 20,670 ﬂoats de-

scribing the entire body. The goal of SMPL is to de-

ﬁne a function M(

β,

θ), where β stands for the shape

parameters and θ for the pose parameters. The func-

tion learns to map the input parameters to human 3D

body meshes (as mentioned above) to ensure that the

output represents valid, realistic body conﬁgurations.

This ability stems from the model’s ability to capture

variations in shape and pose, which enables the pa-

rameters. And this is the ﬁnal formula for the SMPL

model:

θ,

β) = W







(

β,

θ)

| {z }

Deformed template mesh

θ)

|{z}

Joints

, W

|{z}

Skinning weights







(1)

where

(

β,

θ) =

T + B

(

β)

| {z }

Shape deformation

+ B

(

θ)

| {z }

Pose deformation

•

T - the average template mesh;

• J(

θ) - the 3D positions of the skeletal joints that

control posture, based on the pose parameters

θ;

• W - parameters learned from data (training was

performed on 1786 3D scans of humans in differ-

ent poses)

• W - Linear Blend Skinning (LBS) function used

to deform the mesh (the 3D human body) based

on joint rotations and skinning weights.

To improve the usability of the SMPL model, an

important task is to extract 3D meshes from images

without additional inputs such as camera parameters

or pose data. The method used in our work is based

on SMPLify (Bogo et al., 2016), which combines

2D joint detections (2D pose) obtained with methods

such as OpenPose (Cao et al., 2019) with the output

of SMPL. The goal is to match the 3D joints gener-

ated by the SMPL model with the 2D pose estimates

by minimizing an objective function. One of the

challenges is self-penetration, where body parts over-

lap. SMPLify addresses this problem by introducing

a penalty for self-intersection that prevents unrealis-

tic overlaps. Another challenge, depth ambiguity, is

solved using pose priors from the SMPL training data

that penalize implausible poses to improve the accu-

racy of the estimation. An example from our appli-

cation is shown in Figure 2. Further details on the

implementation can be found in (Bogo et al., 2016).

The next tool used is SMPL-X (Pavlakos et al.,

2019), which not only models the entire human body

in 3D, but also facial expressions and hand move-

ments. These two areas are crucial for human com-

munication, and SMPL-X increases realism by in-

cluding detailed expressions for both. To capture the

details of the face and hand, the authors used special-

ized datasets and models:

• a model to represent head meshes. It uses FLAME

(Bolkart and Wuhrer, 2021), which is based on

3,800 head scans.

Generative AI for Human 3D Body Emotions: A Dataset and Baseline Methods

647

• a model to represent the hand. The model MANO

(Romero et al., 2017) is used, which is based on

1,500 hand scans.

Similar to SMPL, principal component analysis

(PCA) was applied to the above datasets to extract

the principal components. Additionally, the num-

ber of joints in the model was increased from 23 to

54 to account for the added complexity of the head

and hand networks. SMPL-X distinguishes between

male, female and, where necessary, gender-neutral

body shapes.

SMPLify-X (Pavlakos et al., 2019) introduces sev-

eral improvements, including VPoser (a variational

autoencoder that learns a distribution over likely hu-

man poses), a reﬁned interpenetration penalty, an im-

proved gender detector, and an overall faster imple-

mentation.

3.2 Generative Model

The purpose of this model is to move values in the va-

riety of open inputs to generate diverse and high qual-

ity postures. In this sense, VPoser (Pavlakos et al.,

2019) compresses body poses into a low-dimensional

latent space and reconstructs valid poses from this

space, ensuring alignment between the joints of the

3D model and the 2D joints recognized by OpenPose.

This is achieved by minimizing an objective function

that relates the two sets of joints and optimizes the

pose accuracy of the model.

VPoser is a deep learning-based body pose prior

designed to model and regularize human 3D poses for

various applications such as animation, virtual reality,

robotics and medicine. Based on a variational autoen-

coder (VAE), VPoser learns a latent space represen-

tation of human postures that enables both pose syn-

thesis and probabilistic inference. The latent space

encodes the most important features of body move-

ments and postures and ensures that the generated or

reconstructed poses are realistic and correspond to the

natural movement constraints of humans. One of the

main advantages of VPoser is its ability to provide

smooth and consistent prioritization for pose gener-

ation, making it particularly effective at minimizing

physically implausible poses.

VPoser is trained using an extensive dataset of hu-

man poses that includes various sources of motion

capture data to ensure diversity and coverage of a

wide range of human movements. In particular, the

training data includes motion capture poses from pub-

licly available datasets, including the CMU motion

capture database Human3.6M (Ionescu et al., 2014)

and the PosePrior dataset (Akhter and Black, 2015).

These datasets are processed using MoSh (Motion

and Shape capture ) (Loper et al., 2014), a technique

that extracts pose parameters in SMPL model format

to ensure compatibility with 3D body shape models.

The resulting pose parameters are then represented in

the form of rotation matrices, which are commonly

used in computer graphics and machine learning for

accurate and smooth rotational transformations. The

VPoser model was trained on approximately one mil-

lion pose samples, with a separate test set of 65,000

poses used to evaluate generalization performance.

These poses are represented in the form of rotation

matrices to ensure consistency with the SMPL body

model.

4 METHODS

This section ﬁrst describes the process of creating the

dataset. It then presents the pipeline used to extract

each frame of the movie, understand it from a skeletal

perspective, and apply body meshes to capture each

person. Finally, details are presented on the genera-

tive model that can be used to create a full 3D body

expression starting from any given human mesh ob-

ject.

4.1 Data Collection

A novel dataset was collected with the help of stu-

dents and professionals at University of Bucharest.

Each participant had to record the expression of dif-

ferent emotions in a place of their choice. The dataset

consists of short videos (about 8 seconds each) in

which the participants express one of six primary

emotions: anger, happiness, surprise, disgust, sad-

ness, and fear. Each video captures the progression of

the emotion, starting from a neutral state, through the

full expression of the emotion, to the return to neu-

trality, providing a comprehensive representation of

the emotional dynamics. For each person and each

emotion, 10 videos were recorded showing the entire

body. To ensure optimal mesh extraction for generat-

ing human-like behaviors for agents, the videos had to

meet several technical requirements. These included a

minimum resolution of FullHD and a simple, prefer-

ably monochrome background (e.g. white). These

conditions were set to avoid complications such as

blurred images or unclear contours that could affect

the accuracy of the shape extraction.

A key advantage of this dataset is that it focuses

on capturing a wide range of emotional expressions

directly from real-life videos. By allowing people to

express their emotions in their own unique way, the

dataset provides a rich diversity of emotional behav-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

648

Figure 1: Snapshots of people and reactions from our data

set. In order of emotions shown such as happy, surprised,

neutral and scared.

ior, which results in the embedding later being highly

expressive and representative of real human reactions.

This diversity is essential for creating realistic 3D

poses and helps ensure that the dataset captures a wide

range of human emotions, which is crucial for appli-

cations in games, animations and virtual reality envi-

ronments.

By extracting SMPL-X objects from video

frames, the dataset provides a variety of poses that

accurately represent the corresponding emotional ex-

pression. This variety is critical for future applica-

tions, especially in production environments such as

the Unreal Engine, where the ability to recreate realis-

tic, human-like behavior for 3D characters is critical

to simulation realism and user perception. By link-

ing different poses - from neutral expressions to emo-

tional state to return to neutrality - this dataset facil-

itates the development of dynamic, believable char-

acter animations that can mimic real-life emotional

responses, signiﬁcantly improving immersion and en-

gagement.

4.2 Extracting Meshes

Once the samples have been collected, the ﬁrst goal

is to extract the 3D meshes representing the different

poses. We chose SMPLify-X (Pavlakos et al., 2019)

as the primary framework because of its ability to cap-

ture body, face and hand expressions. More speciﬁ-

cally, the implementation of SMPLpix and SMPLify-

X (Prokudin et al., 2021) was reused and adapted

it for our dataset. A concrete representation of this

pipeline can be found in Figure 2. In this pipeline, af-

ter experimentation, we created a middle step to rec-

ognize the skeleton from the image frame using the

OpenPose (Cao et al., 2019) (Simon et al., 2017) (Cao

et al., 2017) (Wei et al., 2016) solution, as it was able

to capture not only the skeleton but also the intricate

details of the hands and face that we further required

for our goals. In addition to the image sequence from

the video, the input for this pipeline optionally in-

cludes the gender speciﬁcation. Once the process is

complete, three key ﬁles are output:

• an augmented image showing the mesh and skele-

ton.

• the 3D mesh structure, an .obj ﬁle.

• additional mesh data for further analysis and pro-

cessing, a ﬁle archived in a disk ﬁle.

4.3 Generation of New Poses

When creating realistic, human-like 3D mesh repre-

sentations, one of the ﬁnal steps is to generate poses

and then condition them on a speciﬁc body and adapt

them to certain categories and parameters, Figure

3. To achieve this goal, a Variational Autoencoder

(VAE) based method proved to be the most suitable

approach from the evaluation.

Initially, the experiments started with training a

VAE from scratch. However, the lack of sufﬁcient

training data was a major obstacle to generating high

quality results. Extracting a single 3D mesh is very

computationally intensive, making it impractical to

generate a large enough dataset to train the model ef-

fectively. Given these limitations, we experimented

with transfer learning of VPoser (Pavlakos et al.,

2019), a VAE speciﬁcally designed for the SMPL

model and already trained on a large amount of data.

We reused the basic ideas of the VPoser method to

input pre-generated 3D meshes into the VAE and cre-

ate similar mesh variants. The original mesh was en-

coded into the latent space of the VAE, where trans-

formations were applied to the latent vectors. In par-

ticular, small perturbations were added. The mod-

iﬁed latent vectors were then decoded to generate

new similar meshes. By controlling these transforma-

tions, variations of the original mesh were generated

while preserving its main structural properties. How-

ever, we adapted the original implementation to rep-

resent a Gaussian model for the difference between

two consecutive poses based on the frame and the

Generative AI for Human 3D Body Emotions: A Dataset and Baseline Methods

649

Figure 2: Example for the application of the pipeline to the data set: OpenPose - SMPLify-X. From left to right, OpenPose

is ﬁrst applied to each image extracted from each video in the dataset. The resulting image in the center contains the person

with the 3D skeleton on top. With SMPLify-X, we get the full 3D body skeleton in the right part.

previous pose. This model was trained with the pro-

posed dataset. A pseudocode of our modiﬁed version

of this process is shown in Listing 1. The function

getRandomPoseDi f f (t) extracts from the Gaussian

model mentioned above, constrained by the previous

frame and the time of the sequence.

Listing 1: Sample Python Code.

d e f g e n e r a t e p o s e ( i n i t i a l p o s e ,

num fr a mes ) :

# t h e n o i s e s c a l e

ns = 0 . 5

P o s e {0} = i n i t i a l p o s e

f o r t i n 1 . . n u m f ra m es :

= g e t R andom P o s e D i ff ( t , Pose

t−1

)

Pose

= Pose

t−1

+ n s

In Listing 1, the noise scale represents the degree

of variation introduced into the system and controls

the extent to which the output deviates from the origi-

nal. In particular, it indicates the extent of the change

after the decoding process. In combination with the

original posture parameters, this variation results in

the generation of a modiﬁed version of the mesh rep-

resented by the associated parameters. This process

allows the controlled exploration of different mesh

conﬁgurations while maintaining a relationship to the

original pose.

5 EVALUATION

In this section, we evaluate our work from dif-

ferent perspectives, to which we have mainly con-

tributed. We start with the obtained dataset, eval-

uate the pipeline with the mesh extraction and then

the generative model. Finally, we present an ablation

study from our experiments with other models, dis-

cuss the limitations, observed artifacts and the gen-

eral applicability of our work. The computational re-

sources required for training and inference are also

discussed.

Figure 3: A generative process in action in our framework.

Left: the initial mesh; center: the intermediate result mesh.

Right: the ﬁnal generated pose.

5.1 Dataset: Metrics and Challenges

The Table 1 contains general data about how the

dataset is organized and structured.

Table 1: Dataset overview.

value

participants 28

female participants 6

male participants 22

expressions 6

videos per expression 9 - 10

total number of videos 1677

seconds per video 6 - 10

One of the biggest challenges in creating a new

dataset is the manual work that needs to be done af-

ter the initial recordings. In our case, people have

mimicked emotions that look more like different la-

bels than the ones we were originally looking for, e.g.

anxiety. We decided to isolate these to ensure that

we only considered good quality data. In addition,

there were subjective interpretations of emotion ex-

pressions, such as people expressing fear instead of

scared, which raised questions about the subtle differ-

ences between these two emotions (fear is an endur-

ing emotion, while anxiety is often a more immediate

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

650

reaction). Some videos also had technical limitations,

such as poor quality or problematic viewing angles

that could make it difﬁcult to extract meshes. Despite

these issues, most of the submitted videos were re-

tained and provide a solid basis for creating expres-

sive 3D models from real expressions of emotion.

5.2 Evaluation of the Mesh Extraction

Our experiments have shown that the pipeline dis-

cussed in Section 4 accurately captures the overall

body shape as well as detailed poses for the body, face

and hands (with the hand being particularly well cap-

tured, as shown in 2).

Table 2: Quantitative comparison of the average perfor-

mance of SMPLify-X on different subsets of the dataset and

on the whole dataset. V2V (vertex-to-vertex) error refers to

the error metric used to evaluate how well the estimated hu-

man 3D shape matches the ground truth.

Dataset avg v2v error

anger samples 53.2

happiness samples 54.1

surprise samples 51.4

disgust samples 53.4

sadness samples 52.3

fear samples 50.9

female models 52.9

male models 52.5

Overall 52.6

The results obtained above in 2 conﬁrm that

the proposed dataset and framework have success-

fully achieved their goal of extracting expressive 3D

meshes from custom videos. These extracted meshes

can subsequently be used to generate similar poses

and used for practical applications such as NPC train-

ing or virtual human representations in real-world

scenarios.

5.3 Generative Model Evaluation

Based on our experiments, the generative model

demonstrated the ability to generate high-quality out-

puts with valid and realistic poses. When evaluated

using the vertex-to-vertex (V2V) error metric, the

generated poses showed a strong correspondence to

the original input meshes, indicating minimal devi-

ation in vertex positions. This quantitative assess-

ment conﬁrms that the outputs produced by genera-

tive model are not only plausible in terms of pose but

also sufﬁciently close to the original meshes, validat-

ing the model’s performance in generating accurate

and reliable results.

Figure 4 illustrates the relationship between

Figure 4: V2V similarity decreases as noise levels increase,

indicating greater divergence between the generated and

original meshes due to larger deviations in vertex positions.

vertex-to-vertex similarity (V2V) and increasing

noise applied to the generated mesh. As the noise

level that perturbs the vertex positions increases, the

V2V similarity metric decreases signiﬁcantly. This is

because higher noise leads to larger deviations from

the original mesh, which in turn leads to larger Eu-

clidean distances between the corresponding vertices

of the two meshes. Consequently, as the noise in-

creases, the generated mesh deviates more and more

from the original, leading to a progressive reduction

in overall similarity. The spikes (either up or down)

are caused by the randomness that occurs when new

deviations are generated. In this way, it can be shown

that the generative model was able to produce mean-

ingful poses, which could then be linked to speciﬁc

emotional categories or other criteria such as posture

or movement.

5.4 Artifacts

Overall, the extracted meshes are even if the images

are of good quality, there are some poses that can pro-

duce unwanted artifacts.

In Figure 5, where OpenPose successfully ex-

tracted the hand position in front of the face, we ob-

serve a discrepancy in the application of the frame-

work: the resulting mesh seems to place the hand be-

hind the head. To our understanding, this kind of mis-

alignment is probably due to the function responsible

for aligning the pose generated by the Variational Au-

toencoder (VPoser) with the pose extracted by Open-

Pose. The misalignment could potentially be ﬁxed

by retraining the entire model on the dataset used to

ensure better synchronization between the generated

and extracted poses. However, retraining the compo-

nents of the pipeline from scratch instead of doing

transfer learning would require a signiﬁcant amount

Generative AI for Human 3D Body Emotions: A Dataset and Baseline Methods

651

Figure 5: Artifact in the pipeline: the mesh (right) seems to place the hand behind the head, in contrast to the real image on

the left, where the hands are in front of the head. The skeleton detection phase (center) is correct.

of computational resources.

5.5 Comparison with Other Models

For comparison, we also explored mesh extraction

with ScoreHMR by implementing the code provided

by the authors in (Stathopoulos et al., 2024). In

our initial observations, we noticed a signiﬁcant im-

provement in the speed of mesh extraction from

images compared to SMPLify-X. In particular, in

tests with demo images containing multiple subjects,

ScoreHMR processed the images much faster (about

5x faster for an image with multiple subjects). This

increase in speed is due to the use of diffusion mod-

els that analyze the entire image at once, as op-

posed to classical optimization methods that extract

meshes sequentially. However, a notable drawback of

ScoreHMR is that it sometimes fails to generate the

correct pose (as in Figure 6). Also, it currently only

supports the SMPL model, meaning it cannot extract

face or hand poses, which limits its overall versatility.

Figure 6: 3D mesh, extracted with ScoreHMR. Although

the mesh was extracted in less time compared to SMPLify-

X, it cannot reproduce the facial and hand expressions and

still outputs the position of the hands incorrectly.

5.6 Discussion of Applicability

In our experiments, the computational performance of

the proposed pipeline and our generative model was

evaluated to estimate the time required for mesh ex-

traction and pose generation. With an A100 GPU,

the pipeline processes a single frame in about 52 sec-

onds. For a short video of 10 seconds duration and 30

frames per second, the extraction of the correspond-

ing meshes would therefore take about 4 hours and 20

minutes. Given the size of our dataset, the total time

to process all frames is estimated to be about 7267

GPU hours (or 302 days and 19 hours for a single

GPU), based on 1. However, the processing scales al-

most optimally with the number of GPUs used. Fur-

thermore, the generative model generates meshes at a

rate of 2 seconds, which means that creating N frames

of an animation would take approximately Nx2 sec-

onds. A concrete example: Creating a new animation

of 10 seconds at a rate of 30 frames per second re-

quires ∼ 10 minutes.

Ofﬂine creation of emotion animations is impor-

tant for the industry as it eliminates real-time over-

head and is equivalent to manually created anima-

tions. For example, a game developer can use the

proposed model to create and save animations, which

signiﬁcantly reduces costs and simpliﬁes asset man-

agement and editing. In addition, the full 3D mesh on

each frame provides the necessary detail for the inte-

gration of skin, shaders and clothing, resulting in fully

renderable characters.

6 CONCLUSIONS

This research deals with the creation of 3D anima-

tions that represent human emotions, which are es-

sential in areas such as games, ﬁlm, healthcare, virtual

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

652

reality and robotics. An open-source dataset of vari-

ous emotional expressions was developed and a pro-

cessing pipeline was implemented to analyze skele-

tal and 3D body representations. A generative model

based on Variational Autoencoders (VAEs), in partic-

ular VPoser, was used to generate new 3D poses that

retain emotional nuances. Future work includes inte-

grating these poses into NPC animation pipelines, ex-

tending the dataset for better visualization, and evalu-

ating the impact on user experience in real-world ap-

plications.

ACKNOWLEDGEMENTS

This research is partially supported by the project

“Romanian Hub for Artiﬁcial Intelligence - HRIA”,

Smart Growth, Digitization and Financial Instru-

ments Program, 2021-2027, MySMIS no. 334906

and a grant of the Ministry of Research, Innovation

and Digitization, CNCS/CCCDI-UEFISCDI, project

no. PN-IV-P8-8.1-PRE-HE-ORG-2023-0081, within

PNCDI IV.

REFERENCES

Akhter, I. and Black, M. J. (2015). Pose-conditioned joint

angle limits for 3D human pose reconstruction. In In

proceedings of IEEE CVPR 2015.

Anguelov, D. et al. (2005). Scape: shape comple-

tion and animation of people. ACM Trans. Graph.,

24(3):408–416.

Bogo, F. et al. (2016). Keep it SMPL: Automatic estima-

tion of 3D human pose and shape from a single im-

age. In Computer Vision – ECCV 2016, Lecture Notes

in Computer Science. Springer International Publish-

ing.

Bolkart, T. and Wuhrer, S. (2021). FLAME: A 3d mor-

phable model of the head and face based on 3d scans.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 43(11):2808–2821.

Cao, Z. et al. (2019). Openpose: Realtime multi-person 2d

pose estimation using part afﬁnity ﬁelds. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence.

Cao, Z., Simon, T., Wei, S.-E., and Sheikh, Y. (2017). Real-

time multi-person 2d pose estimation using part afﬁn-

ity ﬁelds. In CVPR.

Hirshberg, D. et al. (2012). Coregistration: Simultane-

ous alignment and modeling of articulated 3d shape.

7577:242–255.

Ionescu, C. et al. (2014). Human3.6m: Large scale datasets

and predictive methods for 3d human sensing in natu-

ral environments. IEEE Transactions on Pattern Anal-

ysis and Machine Intelligence, 36(7):1325–1339.

Kolotouros, N. et al. (2019). Learning to reconstruct 3d

human pose and shape via model-ﬁtting in the loop.

Li, C. et al. (2023). The good, the bad, and why: Unveiling

emotions in generative ai. In ICML 2024.

Loper, M. et al. (2015). Smpl: a skinned multi-person linear

model. ACM Trans. Graph., 34(6).

Loper, M., Mahmood, N., and Black, M. J. (2014). Mosh:

motion and shape capture from sparse markers. ACM

Trans. Graph., 33(6).

Mollahosseini, A., Hassani, B., and Mahoor, M. H. (2017).

Affectnet: A database for facial expression, va-

lence, and arousal computing in the wild. CoRR,

abs/1708.03985.

Park, J. S. et al. (2023). Generative agents: Interactive sim-

ulacra of human behavior. In In the 36th Annual ACM

Symposium on User Interface Software and Technol-

ogy (UIST ’23), UIST ’23, New York, NY, USA. As-

sociation for Computing Machinery.

Pavlakos, G. et al. (2019). Expressive body capture: 3d

hands, face, and body from a single image. In Pro-

ceedings IEEE Conf. on Computer Vision and Pattern

Recognition (CVPR).

Pons-Moll, G. et al. (2015). Dyna: A model of dynamic

human shape in motion.

Prokudin, S., Black, M. J., and Romero, J. (2021). Smplpix:

Neural avatars from 3d human models. In Proceedings

of the IEEE/CVF WACV, pages 1810–1819.

Romero, J., Tzionas, D., and Black, M. J. (2017). Embod-

ied hands: Modeling and capturing hands and bod-

ies together. ACM Transactions on Graphics (TOG),

36(6):245:1–245:17.

Simon, T. et al. (2017). Hand keypoint detection in single

images using multiview bootstrapping. In CVPR.

Stathopoulos, A., Han, L., and Metaxas, D. (2024). Score-

guided diffusion for 3d human recovery. In CVPR.

Sun, J. J. et al. (2021). Eev: A large-scale dataset for study-

ing evoked expressions from video. arXiv preprint

arXiv:2001.05488.

Vrajitoru, D. (2006). Npcs and chatterbots with personality

and emotional response. pages 142 – 147.

Wei, S.-E. et al. (2016). Convolutional pose machines. In

CVPR.

Zadeh, A. et al. (2018a). Multimodal sentiment analysis of

videos: Facial expressions, text, and audio. In Pro-

ceedings of the 2018 Conference on Empirical Meth-

ods in Natural Language Processing (EMNLP), pages

5006–5015. Association for Computational Linguis-

tics.

Zadeh, A. et al. (2018b). Multimodal sentiment analysis of

videos: Facial expressions, text, and audio. In Pro-

ceedings of the 2018 Conference on Empirical Meth-

ods in Natural Language Processing (EMNLP), pages

5006–5015. Association for Computational Linguis-

tics.

Zhao, W. et al. (2024). Open-pose 3d zero-shot learning:

Benchmark and challenges.

Generative AI for Human 3D Body Emotions: A Dataset and Baseline Methods

653