The image shows a prominent research poster displayed at a conference, showcasing work titled "Construction via Diffusion," as part of the "Highlight" section. The poster is affiliated with institutions ETH Zurich, the Computer Vision and Learning Group (VLG), and Meta AI, as indicated by the logos. It is positioned at booth number 182.

The poster presents detailed research on reconstructing global motion and local body pose through a diffusion-based approach. The academic work is credited to authors whose names appear at the top of the poster. Visual diagrams and equations outline the methodology and experimental results, comparing different methods and showcasing significant improvements in motion plausibility and accuracy without foot skating errors. Illustrations and comparative motion reconstructions highlight the effectiveness of their proposed method, which is claimed to be 30 times faster than the previous state-of-the-art method HuMoR during inference.

Experiment results on synthetic and real datasets are displayed with accompanying figures and metrics demonstrating the efficiency and improved performance of their approach. The environment is well-lit and appears to be a professional academic event, possibly a symposium or a conference, with multiple research posters around.
Text transcribed from the image:
Highlight
nstruction via Diffusion ETH Zürich 00 Meta
adlecek?, Siyu Tang', Federica Bogo²
fusing Global and Local Motion
global trajectory
local body pose
Ro=DR(Rt, t, CR)
(Ro, Po) = Dp((Ro, Pt), t, cp)
Inference iteration = 1
MPOP
MROR
Pt
PoseNet
Pov
康
VLG
Training on: AMASS
Computer Vision
and Learning
Croup
CVPR
SEATTLE, WA
Experiments
Test on: AMASS (synthetic noise + occlusions), PROX (RGBD/RGB), EgoBody (RGB)
Evaluation metrics:
Accuracy: MPJPE
Physical plausibility: acceleration + foot skating + foot-floor penetration
46.96
Method
R
R₁
GMPJPE
-vis -occ -all
VPoser-t 33.0 242.6 109.2 0.219
HuMor [67] 42.4 167.9 88.0 0.68 0.230
MDM++ 36.2 71.9 49.2 0.94 0.102
Ours 21.8 57.4 34.8 0.95 0.078
Contt Skat
RGB-D
RGB
1.8
1.9
Method Skating Accel Dist! Skating! Accel Dist
LEMO [100] 0.176
HuMoR [67] 0.117
PhaseMP [72]
Ours
0.038 1.8
34.22
54.76 0.139 23 35.41
0.180 1.8
3.36 0.116 22 9.73
TrajNet
Results on PROX:
No trajectory-pose correlation
→ foot skating
Results on AMASS:
182
>30% improvement over accuracy >67% (RGB-D)/>17% (RGB) improvement over foot skating
上海科技大学
University of
Zurich
Motivation
An N-Poir
e events gented by a 10 le and pr mdp fron Mi
cover partal Inear velochy and I same with a
but inear soles Applinga soliny w
for these part stationer
Contributions
1. Al ser bra ut semide
tens, this her the goonie
2.A10of angle bet ne praneration the ingred
unec ability ting der
14Monctond precin and stone de
anted by the ne
Agonyingined each se
What is an event camera?
Messaan of andra briges durs et
Advantages high temporal restored to b
power consumption, hippie ban
Multiple Solutions
The proposed soler tuft of
One my respond to floping ang
The second compond to fping the edito
Dandipue by checking the
Characte
ling Global Motion Reconstruction
ajNet with local body pose
global motion at inference time
Inference iteration > 1
TrajControl
Pt
PoseNet
Po
ME
Po
R-
TrajNet
RGB
1114
Skating
0.116
0.165
||Accel↓
→ TrajControl improves motion plausibility
2.2
Input
2.7
on PROX dataset
HUMOR
Ours
GT
HMR
30x times faster than HuMoR during inference!