The image showcases a detailed research poster presented at the Conference on Computer Vision and Pattern Recognition (CVPR) held from June 18-22, 2024. The poster, identified with number 168, is titled "Modal Lifelong Navigation." It appears to involve substantial work from multiple contributors, including authors whose names are partly visible. The primary focus of the research is the use of image simulator metadata and large language models to generate language goals for navigation tasks. 

Key findings are illustrated through various graphs and charts, emphasizing efficiency improvements in different navigation methods over time and sensitivity to noise in goal observations. The research demonstrates that skill chaining achieves state-of-the-art (SOTA) results on success rate and SysRep-like Modular Goal-Oriented Navigation (Modular GQN) achieves SOTA on SPL (Success per Length). A selection of process diagrams and graphs visualizes the pipeline for achieving navigation goals, including object goal localization and language goal settings. A small QR code on the top right corner links to further details, perhaps a digital copy of the poster or supplementary materials. The neighboring poster and part of the surrounding conference environment provide context, indicating an academic setting with other attendees visible in the background.
Text transcribed from the image:
Modal Lifelong Navigation
1, Theophile Gervet?, Matthew Chang³, Zsolt Kira,
pozbeh Mottaghi5
age simulator metadata and large language
odels (LLM) to generate language goals
40
30
28.3
+6.6%
LECTING BEST VIEWPOINT
BBOX INFO
20
19.7
Max frame coverage
16.5
10
bed
0
pillow
Modular GOAT
ed." →
BLIP V2
+ prompt
V
"a large bed with a
floral comforter"
>>
with a floral comforter and a pillow in the middle."
ge goal generation pipeline
OBJ
LANG
IMG
GOAL
GOAL
GOAL
CLIP
FEATURE
MATCHING
SuperGLUE
KEYPOINT
MATCHING
ACTION
Success
Success
168
Success
+4.0%
CVPR
JUNE 17-21, 2024
Results
32.3
SPL
15.9
Skill chaining achieves SOTA
on success rate
Modular GOAT achieves SOTA
on SPL
m.h.
102
Modular
CLIP on Wheels
13.1
SenseAct-NN
Skill Chain
88
Sersect-N
Monolithic
Efficiency of SenseAct-NN and Modular method improves over time
Looking for full-time roles!
Habita
G
Georgia SFU
Tech
SOREN FRASER
ENVERSTY
An Analysis of 30
Stanford
00 Meta Al
Maul Kha
Dhn B
What is HSSD?
A dataset of 211 realistic 30 scenes consisting of secretions of houses
modeled using a diverse set of 18,556 unique, high-quality 30 models of
real objects across 466 semantic categories
211 scenes
categories
Modular GOAT
-Modular GOAT
SenseAct-NN Monolithic
-SenseAct-NN Monolithic
32
20
15
24
Efficiency of
navigation improves
10
for both modular and
16
8
5
0
0
2
3 4
5-10
2
3
4 5-10
Number of sub-tasks
Number of sub-tasks
With memory
Without memory
With memory Without memory
32
18
15
end-to-end trained
methods
End-to-end
methods do not
show drop in
performance
26
24
21
10
10
when long-term
17
16
15
8
0
0
Modular GOAT SenseAct-NIN Monolithic
Modular GOAT SenseAct-NN Monolithic
memory is
disabled
Modular methods are more sensitive to noise in goal observations
Object goal
Language goal
Image goal
INSTANCE
MAP
LOCAL POLICY
Without noise With noise
I Without no
it noise With noise
Without rose win noise
50
22
(FMM PLANNER)
32
37.5
16.5
AMIC INSTANCE
MAPPING
GOAL LOCALIZATION
Builds explicit map of the
environment in combination
with path planning for
navigation
Success
8
24
0
Modular Skill Chain Monolithic
GOAT
11
25
0
Modular Skill Chain Monolithic
GOAT
Modular Skill Chain Moroine
GOAT
End-to-end trained methods are robust to noise in goal specification
Semantic annotations
Category hierarchy
Category word cloud
wall art
Maxill