A conference attendee is closely examining a poster presentation titled "LEGaussians: Language Embedded 3D Gaussians." The poster, displayed on a black frame, is authored by Jin-Chuan Shi, Miao Wang, and others, representing the State Key Laboratory of Virtual Reality Technology and *Beihang University*. The content highlights the integration of language embeddings with 3D Gaussian representations to enhance novel view synthesis in 3D scenes. The background, motivation, and methodology sections are outlined, with specific emphasis on the benefits and challenges of embedding language features in 3D scene representations. The colorful diagrams and illustrations on the poster provide visual insight into their approach. The attendee, sporting a navy blue cap and carrying a black backpack, seems deeply engaged with the material, indicative of an academic or professional interest in virtual reality and language processing technologies presented at the event.
Text transcribed from the image:
北京京
航空
航天大
wwwwww
wwwww
BEIHANG
1952
UNIVERSITY
Background
LEGaussians: Language Embedded 3D Gauss
Jin-Chuan Shi Miao Wang12
1State Key Laboratory of Virtual Reality Technology and
Recent works integrate semantics into 3D scenes for open-vocabulary
querying, enabling interaction with LLMS and supporting applications
like scene editing, VR/AR, autonomous driving, and robotic navigation.
NeRFs and 3DGS enhance novel view synthesis from multi-view
images by capturing geometric and appearance details of 3D scenes.
Language-embedded scene representations based on NeRF heavily
depends on MLPs that are resource-intensive in training and rendering.
Motivation
3D Gaussians
offer efficient and high-quality novel view synthesis
directly embed language features in them leads to
huge memory usage and decreased performance
LEGaussia
Language Embedde
3D Gaussians
semantics
uncertainty
Adaptive Spatial Smoothing
➤ Or
position
semantics
Sigmo
uncertainty
ງ
Multi-view Images
DINO
CLIP
Language-embedded
3D Gaussians
Multi-view Features
We introduce LEGaussians: Language Embed
scene representation for open-vo
mization and rende
nd acc