X-KANeRF [KANeRF-benchmarking]: KAN-based NeRF with Various Basis Functions

Is there any basis function can explain the NeRF formula?!

$$\mathbf{c}, \sigma = F_{\Theta}(\mathbf{x}, \mathbf{d}),$$ where $\mathbf{c}=(r,g,b)$ is RGB color, $\sigma$ is density, $\mathbf{x}$ is 3D position, $\mathbf{d}$ is the direction.

To explore this issue, I used Kolmogorov-Arnold Networks (KAN) with various basis functions to fit the NeRF equation based on nerfstudio.

The code might be a bit COARSE, any suggestions and criticisms are welcome!

X-KAN Models (Here are various KANs!)

TODO	Basis Functions	Mathtype	Acknowledgement
√	B-Spline	$$S_i(x) = a_i + b_i(x - x_i) + c_i(x - x_i)^2 + d_i(x - x_i)^3$$	Efficient-Kan
√	Fourier	$$\phi_k(x) = \sin(2\pi kx), \phi_k(x) = \cos(2\pi kx)$$	FourierKAN
√	Gaussian RBF	$$\phi(x, c) = e^{-\frac{|x - c|^2}{2\sigma^2}}$$	FastKAN
√	Radial Basis Function	$$\phi(x, c) = f(|x - c|)$$	RBFKAN
√	FCN	-	FCN-KAN
√	FCN-Interpolation	-	FCN-KAN
√	1st Chebyshev Polynomials	$$T_n(x) = \cos(n \cos^{-1}(x))$$	ChebyKAN
√	2nd-Chebyshev Polynomials	$$U_n(x) = \frac{\sin((n+1)\cos^{-1}(x))}{\sin(\cos^{-1}(x))}$$	OrthogPolyKANs
√	Jacobi polynomials	$$P_n^{(\alpha, \beta)}(x) = \frac{1}{2^n n!} \frac{d^n}{dx^n} \left[ (1-x)^{\alpha+n} (1+x)^{\beta+n} \right]$$	JacobiKAN
√	Hermite polynomials	$$H_n(x) = (-1)^n e^{x^2} \frac{d^n}{dx^n}(e^{-x^2})$$	OrthogPolyKANs
√	Gegenbauer polynomials	$$C_{n+1}^{(\lambda)}(x) = \frac{2(n+\lambda)}{n+1}x C_n^{(\lambda)}(x) - \frac{(n+2\lambda-1)}{n+1}C_{n-1}^{(\lambda)}(x)$$	OrthogPolyKANs
√	Legendre polynomials	$$P_n(x) = \frac{1}{2^n n!} \frac{d^n}{dx^n} \left( x^2 - 1 \right)^n$$	OrthogPolyKANs
-	Laguerre polynomials	$$L_n(x) = \frac{e^x}{n!} \frac{d^n}{dx^n} \left( x^n e^{-x} \right)$$	OrthogPolyKANs
√	Bessel polynomials	$$J_n(x) = \sum_{k=0}^{\infty} \frac{(-1)^k}{k!(n+k)!} \left( \frac{x}{2} \right)^{2k+n}$$	OrthogPolyKANs
√	Fibonacci polynomials	$$F_n(x) = xF_{n-1}(x) + F_{n-2}(x), \quad \text{for } n \geq 2.$$	OrthogPolyKANs
More and More!!!	-	-	-

Performance Comparision on RTX-3090

• Model Setting -> train_blender.sh

Model	hidden_dim	hidden_dim_color	num_layers	num_layers_color	geo_feat_dim	appearance_embed_dim
Nefacto-MLP-A	32	32	2	2	7	8
Nefacto-MLP-B	8	8	8	8	7	8
Others	8	8	1	1	7	8

• nerf_synthetic: lego / 30k

Model	Layer Params $\downarrow$	Train Rays/Sec $\uparrow$	Train Time $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Nerfacto-MLP-A	9902	~170K	~14m	0.71	32.53	0.968	0.0167
Nerfacto-MLP-B	3382	~165K	~14m	0.75	27.11	0.915	0.0621
Nerfacto-MLP	1118	~190K	~13m	0.99	28.60	0.952	0.0346
BSplines-KAN	8092	~37K	~54 m	0.19	32.33	0.965	0.0174
GRBF-KAN	3748	~115K	~19 m	0.50	32.39	0.967	0.0172
RBF-KAN	3512	~140K	~15m	0.71	32.57	0.966	0.0177
Fourier-KAN	5222	~80K	~25 m	0.42	31.72	0.956	0.0241
FCN-KAN(Iters: 4k)	5184	~4K	~90m	0.02	29.67	0.938	0.0401
FCN-Interpolation-KAN	6912	~52K	~40m	0.21	32.67	0.965	0.0187
1st Chebyshev-KAN	4396	~53K	~40m	0.34	28.56	0.924	0.0523
Jacobi-KAN	3532	~72K	~30m	0.37	27.88	0.915	0.0553
Bessel-KAN	3532	~76K	~28m	0.33	25.79	0.878	0.1156
2nd Chebyshev-KAN	4396	~55K	~39m	0.33	28.53	0.924	0.0500
Fibonacci-KAN	4396	~65K	~32m	0.34	28.30	0.922	0.0521
Gegenbauer-KAN	4396	~53K	~40m	0.32	28.39	0.922	0.0514
Hermite-KAN	4396	~55K	~38m	0.37	27.58	0.913	0.0591
Legendre-KAN	4396	~55K	~38m	0.33	26.64	0.893	0.0986

• 360_v2: garden / 30k, todo

Nerfstudio Installation

# create python env
conda create --name nerfstudio -y python=3.8
conda activate nerfstudio
python -m pip install --upgrade pip

# install torch
pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
conda install -c "nvidia/label/cuda-11.7.1" cuda-toolkit

# install tinycudann
pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

# install nerfstudio
pip install nerfstudio==0.3.4

# pip install torchmetrics==0.11.4

# Tab command
ns-install-cli

# !!! If you use `ns-process-data`, please install this version opencv
pip install opencv-python==4.3.0.36

Run

############# kan_basis_type #############
# mlp, bspline, grbf, rbf, fourier,
# fcn, fcn_inter, chebyshev, jacobi
# bessel, chebyshev2, finonacci, hermite
# legendre, gegenbauer
bash train_blender.sh

Docs

• Universal Approximation Theorem vs. Kolmogorov–Arnold Theorem

Acknowledgement

• KANeRF, A big thank you for this awesome work!

  @Manual{kanerf,
  title = {Hands-On NeRF with KAN},
  author = {Delin Qu, Qizhi Chen},
  year = {2024},
  url = {https://github.com/Tavish9/KANeRF},
  }

• nerfstudio

   @inproceedings{nerfstudio,
   	title = {Nerfstudio: A Modular Framework for Neural Radiance Field Development},
   	author = {
   		Tancik, Matthew and Weber, Ethan and Ng, Evonne and Li, Ruilong and Yi, Brent
   		and Kerr, Justin and Wang, Terrance and Kristoffersen, Alexander and Austin,
   		Jake and Salahi, Kamyar and Ahuja, Abhik and McAllister, David and Kanazawa,
   		Angjoo
   	},
   	year = 2023,
   	booktitle = {ACM SIGGRAPH 2023 Conference Proceedings},
   	series = {SIGGRAPH '23}
   }