Poster #18 titled "3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting" is displayed at an academic conference. The poster presents research from a collaboration between ETH Zürich, the University of Tübingen, VLG, and the Tübingen AI Center. It details a method for creating animatable clothed human avatars from monocular videos using deformable 3D Gaussian splatting. 

The overview section summarizes that the technique allows for zero-shot appearances and non-rigid deformations, providing fast novel views synthesis and monocular input. Several logos, including those of ETH Zürich and CVPR, decorate the poster, hinting at its academic and scientific prestige.

The methods section outlines the integration of rigid human articulation with non-rigid deformations under the 3DGS framework. It mentions the replacement of Splatted Human Meshes with a local deformation module and a separate MLP to decode appearance. An illustration provides a visual explanation of the method.

The experiments section showcases a comparison of generated avatars with various state-of-the-art models, highlighting better quality and faster processing times. The novel view synthesis and novel pose animation sub-sections feature visual results demonstrating the capabilities of the proposed method against other methods such as HumanNeRF and ARAH.

Notably, a QR code is present for more information, leading to the project page, and detailed references are provided at the bottom of the poster. A diverse group of attendees can be seen engaging with the research, indicating interest and interaction at this bustling conference venue.
Text transcribed from the image:
17
Z
ETH Zürich
VLG 10
Overview
EBERHARD KARLS
UNIVERSITAT
TUBINGEN
Tübingen Al Center
3DGS-Avatar: Animatable Avatars via Deformable 3D Gaussian Splatting
Zhiyin Qian, Shaofei Wang 1,2,3, Marko Mihajlovic', Andreas Geiger2,3, Siyu Tang¹
1ETH Zürich, 2University of Tübingen, 3Tübingen Al Center
https://neuralbodies.github.io/3DGS-Avatar/
MM X N W
Input: a monocular video
30 min.
training
50 FPS
rendering
Output: animatable 3D Gaussian avatar
Novel view synthesis
Novel pose animation
Experiments
Novel View Synthesis
pose-dependent deformation
novel pose animation
fast training
real-time rendering
monocular input
TL;DR: We create animatable clothed human avatars from
monocular videos using 3D Gaussian Splatting [1]
pose-dependent non-rigid deformation
generalize to unseen poses
fast training (less than 30 minutes)
real-time rendering (50+ FPS)
monocular input
Methods
Canonical
10 Gaussians (G)
t
Optimizable SMPL
Parameters,
Non-rigid Deformation
Rigid Transformation
T (Sec 4.1)
Ta (Sec. 4.2)
CVPR
Mi SEATTLE, WA JUNE 17-21, 2024
Project Page
Novel Pose Animation
x
x
X
✓
x X
NeuralBody
HumanNeRF
ARAH
X W
x
Instant-NVR
✓
NX ✓
InstantAvatar
✓
X X
MonoHuman
x
UV-Volumes
X ✓
x
DELIFFAS
✓
3DGS-Avatar (Ours)
GT
Ours
HumanNeRF [4]
Training:
Inference:
30 min.
>8 days
50+ FPS
0.2 FPS
ARAH [5]
8 days
0.1 FPS
Instant-NVR [6]
5 min.
3 FPS
Ours
30 min.
50+ FPS
Latest
Code Z
Direction di
PSNR 30.61 29.08
30.24
30.52 30.91
31.02
Color MLP
Fe, (Sec. 4.3)
SSIM 97.03 96.16
LPIPS 29.58 52.29
96.80
97.09
97.11
97.08
31.73
36.13
28.62
38.40
Ours NB [3] Human NeRF [4] ARAH [5] Ours Instant-NVR [6]
➤ Compared to current state-of-the-arts [4,5], we achieve
comparable or even better rendering quality while being
hundreds of times faster in both training and inference.
➤ Generalize well to out-of-distribution poses.
HumanNeRF [4]
>8 days
0.2 FPS
ARAH [5]
8 days
0.1 FPS
Duff Gaussian
Rasterization (E
(E45)
Observation Space
Rendered Image
-
ΣΣΑΣ ΣΑΣ Σ
As-isometric-as-possible Regularization [2] |
➤Integrate rigid human articulation with non-rigid
deformation field into the 3DGS framework.
Replace spherical harmonics with a local deformation
aware MLP to decode color.
Apply as-isometric-as-possible regularizations to
generate realistic deformation under novel poses.
Ablation on AIAP Regularization
Full model
w/o Lisocov. Lisopos
➤ AIAP regularization enforces consistent
movement of the Gaussians, thus removing
the artifacts on highly articulated poses.
[1] Kerbi et al. 3D Gaussian Splatting for Real-time Radiance Field Rendering. SIGGRAPH, 2023
[2] Killian et al. Geometric Modeling in Shape Space. SIGGRAPH, 2007
[3] Peng et al. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View
Synthesis of Dynamic Humans. CVPR, 2021
(4] Weng et al. HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular Video. CVPR,
2022
[5] Wang et al. ARAH: Animatable Volume Rendering of Articulated Human SDFs. ECCV, 2022
[6] Geng et al. Learning Neural Volumetric Representations of Dynamic Humans in Minutes. CVPR, 2023
18