BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030656Z
LOCATION:Meeting Room S421\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251215T134200
DTEND;TZID=Asia/Hong_Kong:20251215T135300
UID:siggraphasia_SIGGRAPH Asia 2025_sess104_papers_2076@linklings.com
SUMMARY:Lang3D-XL: Language Embedded 3D Gaussians for Large-scale Scenes
DESCRIPTION:Shai Krakovsky (Tel Aviv University), Gal Fiebelman and Sagie 
 Benaim (Hebrew University of Jerusalem), and Hadar Averbuch-Elor (Cornell 
 University)\n\nEmbedding a language field in a 3D representation enables r
 icher semantic understanding of spatial environments by linking geometry w
 ith descriptive meaning. This allows for a more intuitive human-computer i
 nteraction, enabling querying or editing scenes using natural language, an
 d could potentially improve tasks like scene retrieval, navigation, and mu
 ltimodal reasoning. While such capabilities could be transformative, in pa
 rticular for large-scale scenes, we find that recent feature distillation 
 approaches cannot effectively learn over massive Internet data due to chal
 lenges in semantic feature misalignment and inefficiency in memory and run
 time. To this end, we propose a novel approach to address these challenges
 . First, we introduce extremely low-dimensional semantic bottleneck featur
 es as part of the underlying 3D Gaussian representation. These are process
 ed by rendering and passing them through a multi-resolution, feature-based
 , hash encoder. This significantly improves efficiency both in runtime and
  GPU memory. Second, we introduce an Attenuated Downsampler module and pro
 pose several regularizations addressing the semantic misalignment of groun
 d truth 2D features. \nWe evaluate our method on the in-the-wild HolyScene
 s dataset and demonstrate that it surpasses existing approaches in both pe
 rformance and efficiency. Code will be available.\n\nRegistration Category
 : Full Access, Full Access Supporter\n\nSession Chair: Xuejin Chen (Univer
 sity of Science and Technology of China)\n\n
END:VEVENT
END:VCALENDAR