Neural Lumigraph Rendering

[ cvpr  differetial-render  nerf  deep-learning  synthetic  2021  3d-mesh  riren  lumigraph  best-paper  ]

Neural Lumigraph Rendering was accepted for CVPR 2021 Oral and best paper candidate. This paper proposes a method which performs on par with NeRF on view interpolation tasks while providing a high-quality 3D surface that can be directly exported for real-time rendering at test time. code is publically available.

Screen Shot 2021-06-23 at 11.47.47 PM

The major contribution of this paper is that it addresses the speed limitation of neural radiance field method, which is state of art neural rendering method, by adopting an SDF-based sinusoidal representation network (SIREN) as the backbone of our neural rendering system and export a 3D face mesh could be used for real time rendering.

Representation

We express the continuous shapes of a scene as the zero level set \(S_0=\{x\vert S(x)=0\}\) of a signed distance function (SDF) \(S(x;\theta):\mathbb{R}^3\to\mathbb{R}\) where \(x\in\mathbb{R}^3\) is a location in 3D space and \(\theta\) are the learnable parameters of our SIREN-based SDF representation.

RIREN is a MLP network is uses sin function as the activation function instead of ReLU or others.

we model appearance as a spatially varying emission function, or radiance field, E for directions \(r_d\in\mathbb{R}^3\) defined in a global coordinate system. This formulation does not allow for relighting but it enables photorealistic reconstruction of the appearance of a scene under fixed lighting conditions.

Together, the radience field is written as:

\[E(x,r_d,n,\theta,\phi):\mathbb{R}^9\to\mathbb{R}^3\]

here n is surface normal, \(\theta\) is the parameter for shape network and \(\phi\) is the parameter for radiance network.

Neural Rendering

We solve this task in two steps:

  • We find the 3D surface as the zero-level set \(S_0\) closest to the camera origin along each ray;
  • We resolve the appearance by sampling the local radiance E.

To export mesh for rendering, the following two steps are need respectively:

  • we use the marching cubes algorithm to extract a high-resolution surface mesh from the SDF S voxelized at a resolution of 512;
  • To export the appearance, we resample the optimized emissivity function E to synthesize projective textures Ti for N camera poses and corresponding projection matrices.

Training

We supervise our 3D representation using a set of m multi-view 2D images \(I=\mathbb{R}^{m\times w\times h\times 3}\) with known object masks \(M=\mathbb{R}^{m\times w\times h}\) where 1 marks foreground. The following regularization term is used:

  • First, we minimize an L1 image reconstruction error for the true foreground pixels;
  • Second, we regularize the S by an eikonal constraint to enforce its metric properties important for efficient sphere tracing
  • Third, we restrict the coarse shape by enforcing its projected pattern to fall within the boundaries of the object masks
  • linearize the angular behavior using a smoothness term.

We optimize the loss in mini-batches of 50,000 individual rays sampled uniformly across the entire training dataset. We have found a large batch size and uniform ray distribution to be critical to prevent local overfitting of SIREN, especially for the high-frequency function E. We implement the MLPs representing S and E as SIRENs with 5 layers using 256 hidden units each.

Experiment Result

Quantiative evaluation is available as below (ST: sphere-trace, RAS: rasterize):

Screen Shot 2021-06-24 at 12.04.47 AM

Some visual comparisons are available below, which shows that the propose method NLR preserves more details and is more sharp.

Screen Shot 2021-06-24 at 12.05.30 AM

Written on June 24, 2021