SC-NeuS: Consistent Neural Surface Reconstruction
from Sparse and Noisy Views

To enable multi-view consistent constraints, the essential requirement of the geometry cues is that they need locate on the explicit surface, i.e., the zero level set of the signed distance field $f(x,\bm{\theta})$. Considering a 2D feature point $p\in R^{2}$ in the reference image $I_{i}$ with camera pose $T_{i}$, we seek to compute its intersection point $\bm{P}^{*}\in R^{3}$ on the surface geometry of signed distance field $f(x,\bm{\theta})$. According to volume rendering of the signed distance function [22, 35], there exists a ray length value $t^{*}$ such that:

$$\bm{P}^{*}=\bm{c_{i}}+t^{*}\bm{v},\quad f(\bm{P}^{*},\bm{\theta})=0,$$

where $\bm{c_{i}}$ and $\bm{v}$ are the camera center point and casting ray of $p$ respectively.

Although IDR [44] have provided a differentiable intersection derivation for $\bm{P}^{*}$, however, which is somewhat too slow to enable an efficient neural surface learning. Therefore, we propose a new differentiable on-surface intersection for fast neural surface learning. Specifically, as shown in Fig. 3, we first uniformly sample points in the casting ray $\bm{v}$ of 2D feature point $p$ with sampling depth value set $\mathbf{T}=\{t_{k}\}$. Then we find the depth value $t_{k}$ such that $f(\bm{c_{i}}+t_{k}\bm{v},\bm{\theta})f(\bm{c_{i}}+t_{k+1}\bm{v},\bm{\theta})<0$. Finally, we move $t_{k}$ along the casting ray $\bm{v}$ to the on-surface intersection $\bm{P}^{*}(T_{i},\bm{\theta},\bm{v})$ following:

$$\bm{P}^{*}(T_{i},\bm{\theta},\bm{v})=\bm{c_{i}}+t_{k}\bm{v}-\frac{\bm{v}}{\langle\frac{\partial f}{\partial x},\bm{v}\rangle}f(\bm{c_{i}}+t_{k}\bm{v},\bm{\theta}).$$

(1)

Based on our differentiable intersection, we further define effective losses to neural surface learning in the multi-view scenario. Specifically, we utilize two kinds of on-surface geometry cues, i.e., 3D sparse points and patches (Fig. 4), and formulate view-consistent losses for these on-surface geometry cues, including view-consistent re-projection loss and view-consistent patch-warping loss respectively.

View-consistent Re-projection Loss. Considering a pair of 2D feature correspondence ($p_{k}^{i},p_{k}^{j}$) from reference image $I_{i}$ (camera pose $T_{i}$) and target image $I_{j}$ (camera pose $T_{j}$) with $p_{k}^{i}\in I_{i},p_{k}^{j}\in I_{j}$, we compute the on-surface intersection 3D point $\bm{P}_{k}^{ij}$ via our differentiable intersection. By re-projecting $\bm{P}_{k}^{ij}$ back to $I_{i}$ and $I_{j}$, we get the re-projection location as $\bar{p}_{k}^{i}=\pi(\bm{P}_{k}^{ij},T_{i})$, $\bar{p}_{k}^{j}=\pi(\bm{P}_{k}^{ij},T_{j})$, where $\pi(\cdot)$ is the camera projection operator. For a geometry-consistent surface reconstruction, the re-projection error between $p_{k}^{i}\to\bar{p}_{k}^{i}$ and $p_{k}^{j}\to\bar{p}_{k}^{j}$ should be minimized. Then we formulate the view-consistent re-projection loss $L_{r}$ for all of possible sparse correspondence as:

$$L_{r}=\sum_{i,j}\sum_{k\in N_{k}}(|p_{k}^{i}-\pi(\bm{P}_{k}^{ij},T_{i})|+||p_{k}^{j}-\pi(\bm{P}_{k}^{ij},T_{j})|).$$

View-consistent Patch-warping Loss. We also consider the on-surface patch (Fig. 4) to utilize the geometric structure constraints to further improve the neural surface learning. Similar to the patch warping in traditional MVS method [17, 30, 29], we warp the on-surface patch to multi-view images but in a differentiable way using our differentiable multi-view intersection. Specifically, for a small patch $\bm{s}$ on the surface which is observed by image pair $I_{i},I_{j}$, we represent the plane equation of $\bm{s}$ in the camera coordinate of the reference image $I_{i}$ as:

$$\bm{n}^{T}\bm{p}+d=0,$$

where $\bm{p}(T_{i},T_{j})$ is the differentiable multi-view intersection point from $I_{i},I_{j}$ with camera poses $T_{i},T_{j}$, $\bm{n}$ is the normal computed with automatic differentiation of the signed distance filed $f(x,\bm{\theta})$ at $\bm{p}(T_{i},T_{j})$. Suppose that the $\bm{s}$ is projected to $I_{i},I_{j}$ to obtain image patches $\bm{s}_{i}\in I_{i},\bm{s}_{j}\in I_{j}$ respectively, for image pixel $x\in\bm{s}_{i}$ and its corresponding pixel $x^{\prime}\in\bm{s}_{j}$, we have:

$$x=\bm{H}x^{\prime},\\ \bm{H}=K_{i}(R_{i}R_{j}^{t}-\frac{R_{i}(R_{i}^{T}t_{i}-R_{j}^{T}t_{j})\bm{n}^{T}}{d})K_{j}^{-1},$$

where $\bm{H}$ is the homography matrix, $T_{i}=\{R_{i}|t_{i}\},T_{j}=\{R_{j}|t_{j}\}$, $K_{i},K_{j}$ are the intrinsic camera matrix for image pair $I_{i},I_{j}$.

We use the normalization cross correlation (NCC) of patches ($\bm{s}_{i},\bm{s}_{j}$) as the view-consistent patch-warping loss as :

$$L_{ncc}(\bm{s}_{i},\bm{s}_{j})=\frac{Cov(I_{i}(\bm{s}_{i}),I_{j}(\bm{s}_{j}))}{Var(I_{i}(\bm{s}_{i}))Var(I_{j}(\bm{s}_{j}))},$$

where $Cov$ and $Var$ donates the covariance and variance for color identity of patches ($\bm{s}_{i},\bm{s}_{j}$) respectively.

Based on the view-consistent losses, we formulate the objective function $E$ as:

$$E=L_{color}+\lambda_{r}L_{r}+\lambda_{ncc}L_{ncc}+\lambda_{reg}L_{reg},$$

(2)

with $L_{r}$ and $L_{ncc}$ are the view-consistent re-projection loss and patch-warping loss defined above, and $L_{color}$ and $L_{reg}$ are the color rendering loss and Eikonal regularization loss proposed by NeuS [35] as:

$$L_{color}=\frac{1}{N}\sum_{i}^{N}|\mathcal{R}(f(x,\bm{\theta}),c(x,\bm{\theta_{c}},\bm{v}),T_{i})-I_{i}|,$$

$$L_{reg}=\frac{1}{M}\sum_{i}^{M}(||\bigtriangledown f_{\theta}||_{2}-1)^{2},$$

where $\mathcal{R}(f(x,\bm{\theta}),c(x,\bm{\theta_{c}},\bm{v}),T_{i})$ is the volume rendering image from $f(x,\bm{\theta}),c(x,\bm{\theta_{c}},\bm{v})$ to view $T_{i}$.

So in summary, we propose to jointly learn the signed distance filed $f(x,\bm{\theta})$, radian field $c(x,\bm{\theta_{c}},\bm{v})$ and camera poses $T=\{T_{i}\}$ to optimize the objective function $E$ in an end-to-end manner following:

$$\{\bm{\theta^{*}},\bm{\theta_{c}^{*}},T^{*}\}=\arg\min_{\bm{\theta},\bm{\theta_{c}},T}E.$$

(3)

Network Training. In the early stage during the network training, since the signed distance field $f(x,\bm{\theta})$ doesn’t converge very well, we choose a warm-up strategy to assist the convergence. Specifically, for the differentiable on-surface intersection of 3D sparse point, we use the rendering depth information of the recent signed distance field $f(x,\bm{\theta})$ to filter out outlier on-surface point intersection. Given a 2D feature point $p$ and its on-surface intersection point $\bm{P}^{*}$ along casting ray $\bm{v}$, we compute its depth value $t_{d}$ [34] and get its 3D point re-projection $\bm{P}_{d}=\bm{c}_{i}+t_{d}\bm{v}$. If the distance between $\bm{P}^{*}$ and $\bm{P}_{d}$ is larger than a threshold, we set $\bm{P}^{*}=\bm{P}_{d}$ to perform the joint learning. After the warm-up for a certain training epoch, we conduct the learning according to equation 3 until the final convergence for both signed distance field $f(x,\bm{\theta})$ and camera poses $T$.

Coarse-to-Fine Learning. As like BARF [18], we also adopt the similar coarse-to-fine positional encoding learning strategy, for a better convergence during the joint learning of neural surface and camera poses. Please refer to BARF [18] for more details.

Dataset. Similar with previous neural surface reconstruction approaches [12, 44, 35], we choose to evaluate our approach on the public DTU dataset [1] with 15 different object scan. The DTU dataset contains from 49 to 64 images at a resolution of $1200\times 1600$ for each object scan with known camera intrinsic matrix and ground truth camera poses. For sparse views, we follow [19] and  [18] to randomly select as few as 3 views for each object scan, and then synthetically perturb its camera pose with an additive Gaussian noise $\mathcal{N}(0,0.15)$, thus collecting a sparse version of DTU dataset for the subsequent evaluation. Besides, we also evaluate on 7 challenging scenes from low-res set of the BlendedMVS dataset [42] which includes $31-143$ images at a resolution of $768\times 576$. Similar to pre-processing as like our DTU dataset, we also select 3 views from them and obtain their noisy initial poses.

Baselines. We compare our approach with previous state-of-the-art approaches which also perform neural surface reconstruction by joint learning of neural surface and camera poses, including BARF [18]and IDR [44]. Besides, although NeuS [35] doesn’t conduct camera pose optimization, since it is a state-of-the-art neural surface reconstruction approach, we also compare our approach with NeuS by incorporating the coarse-to-fine strategy of BARF [18], called “NeuS-BARF”, to enable a fair comparison. Since IDR use an extra object mask for neural surface learning, for fair comparison of IDR and NeuS-BARF, we additionally conduct an experiment by using extra object mask supervision for NeuS-BARF, named “NeuS-BARF*”, during the subsequent evaluations.

Camera Pose Comparison. Table 1 demonstrates the average RMSE accuracy (including both translation and rotation error) between the estimated camera poses and the ground truth camera poses on DTU dataset, using different comparing approaches, including BARF, IDR, NeuS-BARF, NeuS-BARF* and ours respectively. Among all the comparing approaches, the NeRF-like approach BARF, achieves worse RMSE accuracy than the other approaches. This makes sense since other approaches (including our approach) adopt the signed distance field to represent the object’s geometry, which is more powerful than the radiance filed used in BARF. Although IDR and NeuS-BARF (NeuS-BARF*) achieve various RMSE accuracy in each object scan of DTU dataset respectively, in average they achieve the same level of RMSE accuracy, which means they perform similar camera pose estimation quality.

In contrast, our approach significantly outperforms all the other baselines in the RMSE accuracy (in both the translation and rotation errors) for camera pose estimation. This shows the multi-view consistent constraints used in our SC-NeuS takes effects for camera pose estimation, than the other baseline approaches which performs the camera pose optimization along with the neural surface learning with single-view independent regularization.

Surface Reconstruction Quality. We also compare the surface reconstruction quality between the different comparing approaches. Table 2 demonstrates the quantitative results on Chamfer Distance metric using different approaches evaluated on DTU dataset. Similarly, our approach achieve consistently much better Chamfer Distance accuracy than the other comparing approaches. Fig. 5 illustrates some visual comparison results of the comparing approaches. Even though BARF can achieve acceptable camera pose estimation quality, it still fails to achieve fine surface reconstruction results (see the first row of Fig. 5).

Besides, due to lack of extra mask object supervision, NeuS-BARF can’t achieve accurate enough camera pose estimation and thus fails to reconstruct fine object surface. This demonstrates that coarse-to-fine position embedding proposed in BARF is not effective to sparse view setting, even utilizing the neural signed distance filed representation as like NeuS. In contrast, our approach choose view-consistent constraints to regularize the joint learning of neural surface representation and camera pose, leading to geometry-consistent surface reconstruction with fine-grained details. Please see the fine-grained detail reconstruction by our approach, which is also better than the state-of-the-art neural surface reconstruction approach like NeuS and SparseNeuS, even with ground truth camera poses as input (Fig. 5). Please refer to our supplementary materials for more comparing results.

Except from the DTU dataset, we also perform evaluation on BlendedMVS dataset to see how our approach behave across different kinds of datasets. Fig. 6 shows some visual comparing surface reconstruction results using different comparing approaches, including NeuS-BARF, NeuS-BARF*, IDR and our approach. According to the comparison, our approach can achieve much better surface reconstruction quality with fine-grained details than the other approaches. Here we don’t include BARF for visual comparison, since BARF fails to converge in most of the comparing cases. Please refer to our supplementary materials for more quantitative and qualitative comparing results using different approaches on BlendedMVS dataset.

The re-projection loss and patch-warping loss serve as two main components of our approach. We conduct an ablation study experiment, to see how these two losses take effect on the final quality of both surface reconstruction and camera pose estimation.

View-consistent Re-projection. We first implement a variant version of our full system without using the view-consistent re-projection loss, termed as ’w/o $L_{r}$’, and perform the surface reconstruction on the DTU dataset. Table 3 shows the average RMSE accuracy for camera pose estimation quality and CD accuracy for the surface reconstruction quality for ’w/o $L_{r}$’ and our full system (termed as ’Full’). We can see there are large accuracy decrease for both RMSE and CD between ’w/o $L_{r}$’ and ’Full’ systems. This means the view-consistent re-projection loss serves major contribution in our SC-NeuS for the final geometry-consistent surface reconstruction and accurate camera pose estimation. But please note that ’w/o $L_{r}$’ still outperforms other comparing approaches including BARF, IDR and NeuS-BAFR, by achieving better average RMSE accuracy and CD accuracy in Table 1 and  2.

View-consistent Patch-warping. We also implement a variant system without using the view-consistent patch-warping loss, termed as ’w/o $L_{ncc}$’. According to the average RMSE accuracy and CD accuracy comparison between ’w/o $L_{ncc}$’ and ’Full’ in Table 3, we can see that ’w/o $L_{ncc}$’ also achieve worse accuracy values than ’Full’ in both RMSE for camera pose estimation and CD for surface reconstruction quality, even thougth the quality decreases are not that much compared with those from ’w/o $L_{r}$’ to ’Full’.

Fig. 7 shows the visual comparing surface reconstruction results of two example from DTU dataset, using ’w/o $L_{r}$’, ’w/o $L_{ncc}$’ and ’Full’ respectively. We can see there are certain surface quality decrease for our full system (’Full’) without using the view-consistent re-projection loss (’w/o $L_{r}$’). Even though our approach can achieve fine surface reconstruction without using view-consistent patch-warping loss (see the results of ’w/o $L_{ncc}$’), we can obvious fine-grained details enhancement by adding the view-consistent loss to our full system (see the results of ’Full’). This means that view-consistent patch-warping loss takes more effective for fine-grained details, while view-consistent re-projection loss works better to boost up the joint learning quality of neural surface and camera pose.

Our approach’s first limitation is that influence from the quality of 2D feature point’s matching. Without enough feature point matching in challenging cases like low texture or light changing, our approach couldn’t perform well for nice surface reconstruction results. Large camera poses variation between sparse views would also make our approach failed for feasible joint optimization. In the further, we would like to use more robust explicit surface priors for high reliable neural surface reconstruction.