Neural Gaussian Scale-Space Fields

Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, Thomas Leimkühler

Max-Planck-Institut

SIGGRAPH 2024

https://neural-gaussian-scale-space-fields.mpi-inf.mpg.de/


Abstract

Gaussian scale spaces are a cornerstone of signal representation and processing, with applications in filtering, multiscale analysis, anti-aliasing, and many more. However, obtaining such a scale space is costly and cumbersome, in particular for continuous representations such as neural fields. We present an efficient and lightweight method to learn the fully continuous, anisotropic Gaussian scale space of an arbitrary signal. Based on Fourier feature modulation and Lipschitz bounding, our approach is trained self-supervised, i.e., training does not require any manual filtering. Our neural Gaussian scale-space fields faithfully capture multiscale representations across a broad range of modalities, and support a diverse set of applications. These include images, geometry, light-stage data, texture anti-aliasing, and multiscale optimization.

AI

高斯尺度空间 是信号表示和处理的核心基石,其在过滤、多尺度分析、抗锯齿、以及许多其他应用中的应用广泛。然而,获取这样的一个尺度空间是昂贵且繁琐的,尤其是在连续表示(如神经领域)中,学习一个高斯尺度空间所需的资源和时间远远超过了手动筛选的过程。
本文提出了一种高效的、轻量级的方法来学习一个完全连续、各向异性的高斯尺度空间的模型。基于傅立叶特征模数化和Lipschitz bounding,通过自定义训练任务,即不需要任何人工过滤过程即可完成高斯尺度空间的学习和应用。这种方法具有高度的灵活性和适应性,可以广泛应用于各种信号表示和处理的应用场景中,包括但不限于图像、几何、光谱数据、纹理抗锯齿以及多尺度优化等。


Gaussian scale spaces

Note

对于一个信号f:RdiRdo, 对应的线性高斯尺度空间定义为

fΣ(x):=1(2π)didet(Σ)Rdif(xτ)exp(12τTΣ1τ)dτ

其实也就是f和一个由正定矩阵Σ 定义的高斯核的连续卷积。

Pasted image 20240707141859.png


Position-Encoding

γ(x)=(λ1cos(2πa1Tx)λ1sin(2πa1Tx)λmcos(2πamTx)λmsin(2πamTx)).

Lipschitz continuity

Note

函数的Lipschitz 连续定义了函数变化的快慢,形式上说,存在一个Lipschitz bound c

f(x1)f(x2)pcx1x2p

Pasted image 20240707153514.png


Method

表示

F(x,Σ^)=Ψθ(γ(x,Σ^))λi(Σ^)=exp(aiTΣ^ai).F(x,Σ)=Ψθ(γ(x,h(Σ)))Σ^=h(Σ)μΣ

Lipschitz-bounded function construction

SVD分解 $$W_k = U_kS_KV_K^T$$

||Wk||2q$$$Sk$Sigmoid![](https://cdn.mathpix.com/snip/images/rvQUOUmMN7FL57c6mNv6FVsUF25exoQnpO95oDaaM.original.fullsize.png)

Pasted image 20240707161317.png


![[dragon.mp4]]
![[bbq.mp4]]