Learning Novel View Synthesis from Heterogeneous Low-light Captures

TL;DR

This work proposes to decompose illumination, reflectance, and noise from input views according to that reflectance remains invariant across heterogeneous views and achieves superior visual quality and numerical performance for synthesizing novel views compared to state-of-the-art methods.

Abstract

Neural radiance field has achieved fundamental success in novel view synthesis from input views with the same brightness level captured under fixed normal lighting. Unfortunately, synthesizing novel views remains to be a challenge for input views with heterogeneous brightness level captured under low-light condition. The condition is pretty common in the real world. It causes low-contrast images where details are concealed in the darkness and camera sensor noise significantly degrades the image quality. To tackle this problem, we propose to learn to decompose illumination, reflectance, and noise from input views according to that reflectance remains invariant across heterogeneous views. To cope with heterogeneous brightness and noise levels across multi-views, we learn an illumination embedding and optimize a noise map individually for each view. To allow intuitive editing of the illumination, we design an illumination adjustment module to enable either brightening or darkening of the illumination component. Comprehensive experiments demonstrate that this approach enables effective intrinsic decomposition for low-light multi-view noisy images and achieves superior visual quality and numerical performance for synthesizing novel views compared to state-of-the-art methods.

Figures (7)

• Figure 1: Images rendered from NeRF mildenhall2020nerf trained on the input views with heterogeneous brightness (left) present unevenly illuminated artifacts (middle and right).
• Figure 2: An overview of the proposed framework. In the first stage, our method learns disentangled reflectance, illumination, and noise components from low-light input images with varying brightness based on the Retinex theory. In the second stage, our method can robustly enhance or darken the illumination component. The adjusted image is synthesized by the product of the decomposed reflectance and the adjusted illumination.
• Figure 3: The decomposition results of our approach on the test views of the Shrub and Link scenes. The reconstructions with and without the noise component are also presented. The noise maps are normalized to $[0, 0.4]$ for the visualization.
• Figure 4: View synthesis and enhancement comparisons between our approach and the SCI+NeRF-W, DCE+NeRF-W, and EnGAN+NeRF-W methods on the same test views of four scenes. The reference images are from the high scale and they are pre-processed by a non-local mean denoiser.
• Figure 5: Demonstrations on the generalization ability of our illumination adjustment module on the Link and Plant scenes. Note that the $\epsilon$ values $0.125$, $0.25$, $4$, and $8$ are not observed during the training. The second and fourth rows present the adjusted illumination components. The first and third rows are the product of the reflectance and the adjusted illumination.