LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS
Wanhua Li, Yujie Zhao, Minghan Qin, Yang Liu, Yuanhao Cai, Chuang Gan, Hanspeter Pfister
2025-07-11
Summary
This paper talks about LangSplatV2, a new technology that makes it much faster and more accurate to find things in 3D scenes using just text, without needing a big, slow decoder.
What's the problem?
The old way of connecting words to 3D objects for searching in virtual worlds was too slow and used too much computing power because it needed a heavy decoder, which made it hard to use for real-time tasks.
What's the solution?
The researchers got rid of the slow part by using something called Gaussian Splatting, which spreads information smoothly in space, and they made it even better by having only a few key numbers describe each object's language link. This way, the computer can answer text searches in a 3D world super fast while still being accurate.
Why it matters?
This matters because it lets people or systems quickly search and interact with complex 3D environments using regular words, which opens up possibilities for games, VR, robotics, and more without needing super expensive computers.
Abstract
LangSplatV2 improves the speed and accuracy of 3D open-vocabulary text querying by eliminating the need for a heavyweight decoder through Gaussian Splatting and sparse coefficient splatting.