In-Young Cho, Yuchi Huo, Sung-Eui Yoon, KAIST, ACM Transactions on Graphics (Proc. of SIGGRAPH 2021)
Figure: Visual comparisons between Vanilla models and their manifold counterparts (ours). It is challenging for both image-space and sample-space models to reconstruct fine details caused by high-frequency textures or complex geometries, especially on reflective/refractive objects. Manifold models alleviate these issues by providing reconstruction networks with discriminative path cluster information as auxiliary inputs. Path-space contrastive learning, which uses dense reference labels, leverages rich sample features to distinguish fine details from noise while remedying the sparsity. "My Kitchen" by tokabilitor under CC0. "BATH" by Ndakasha under CC0.
Our manifold learning framework analyzes light path clusters in Monte Carlo path tracing, resulting in better image reconstruction. Our contrastive method fully utilizes sparse high-dimensional auxiliary features.
Monte Carlo (MC) path tracing has been widely used to synthesize realistic images. However, it takes extensive time to render a high-quality image since it requires to sample numerous light paths for each pixel. Hence, MC image reconstruction methods have been actively studied to remove rendering noises and recover clean images.
This study proposes a light path clustering framework to further improve MC image reconstruction. Though image-space features (e.g., surface normal, depth, texture maps) have significantly contributed to MC denoising, direct utilization of high-dimensional light paths has not yet been sufficiently explored. This paper proposes a contrastive manifold learning framework that reduces the dimensionality of path space for MC reconstruction models to exploit path-space features effectively.