LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS

Wanhua Li, Yujie Zhao, Minghan Qin, Yang Liu, Yuanhao Cai, Chuang Gan, Hanspeter Pfister

Keywords: Gaussian Splatting, Real time, Language Field

TL;DR: We present LangSplatV2, which achieves high-dimensional feature splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for high-resolution images.

Abstract:

In this paper, we introduce LangSplatV2, which achieves high-dimensional feature
splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for
high-resolution images, providing a 42 × speedup and a 47 × boost over LangSplat
respectively, along with improved query accuracy. LangSplat employs Gaussian
Splatting to embed 2D CLIP language features into 3D, significantly enhancing
speed and learning a precise 3D language field with SAM semantics. Such advancements in 3D language fields are crucial for applications that require language
interaction within complex scenes. However, LangSplat does not yet achieve real-
time performance (8.2 FPS), even with advanced A100 GPUs, severely limiting
its broader application. In this paper, we first conduct a detailed time analysis of
LangSplat, identifying the heavyweight decoder as the primary speed bottleneck.
Our solution, LangSplatV2 assumes that each Gaussian acts as a sparse code within
a global dictionary, leading to the learning of a 3D sparse coefficient field that
entirely eliminates the need for a heavyweight decoder. By leveraging this sparsity,
we further propose an efficient sparse coefficient splatting method with CUDA optimization, rendering high-dimensional feature maps at high quality while incurring
only the time cost of splatting an ultra-low-dimensional feature. Our experimental
results demonstrate that LangSplatV2 not only achieves better or competitive query
accuracy but is also significantly faster. Codes and demos are available at our project page: https://langsplat-v2.github.io.

Primary Area: Applications (e.g., vision, language, speech and audio, Creative AI)

Submission Number: 16758