Sampling and Rendering

Leader of the group: Gurprit Singh, PhD

Vision and Research Strategy

We are working at the frontiers of sampling, human perception and Monte Carlo rendering. In rendering, Quasi-Monte Carlo (QMC) samplers like Halton and Sobol have shown superior error convergence rate compared to Monte Carlo samplers. Although QMC samplers would result in lower error magnitude, the perception of error (as noise) on rendered images can still be visually displeasing. We are developing theoretical models that bridges the gap between the human visual system and Monte Carlo rendering. We are developing samplers that not only reduce the error magnitude but can give perceptually pleasing error distribution especially at smaller sample count. This research is introducing human perception as an important pillar to Monte Carlo rendering.
Another aspect where sample correlations are far important is in the placement of objects in nature. Various editing tools are developed over the decade that tries to mimic the natural order of things, e.g., the placement of trees and flowers. However, this requires capturing both the local and global correlations for both stationary and non-stationary distributions. Over the past few years, several neural network architectures have been developed to capture these correlations. However, these pipelines are mostly restricted to pixelated data. Our goals extend beyond pixels and deal with data in continuous space.
Our group is also developing modern neural network pipelines for both forward and inverse rendering problems. Since sampling is at the heart of computing, we are also expanding towards improving the fundamentals of neural networks by investigating samplers that can help reduce the generalization error. We hope that our research would not remain limited to computer graphics and vision but may also benefit other branches of computing by establishing a coherent exchange of knowledge.

Research Areas and Achievements

Our research areas are broadly categorized under sampling and rendering. Over the past few years we have published in both categories.

Sampling To design sample correlations we propose a deep learning pipeline that only requires a target power spectrum designed by the user. We use neural network as an advance optimizer. This was published at SIGGRAPH Asia 2019. This has applications from rendering, dithering to object placement in rendering. We also recently demonstrate the impact of sample correlations for deep implicit field-based 3D reconstruction from a single image, which was published in ECCV 2020 for Oral presentation (2% acceptance). Recently, we have shown the impact of sample correlations on data visualization. We will be presenting this work titled Blue noise plots at Eurographics 2021. We have also submitted our recent work on extending sampling beyond traditional approaches to SIGGRAPH 2021.

Rendering Motivated from our deep point correlation design architecture from SIGGRAPH ASIA 2019, we embark on a fabrication project that uses volumetric rendering algorithm for light field 3D printing. We develop a continuous space pipeline that allows optmizing the light field imagery in a full 3D space. Previous methods were mostly limited to optimizing in 2D subspaces (as layers). This work was published at SIGGRAPH Asia 2020. We have also recently formulated a theoretical model that guides the inclusion of human perception into Monte Carlo rendering in a systematic manner. This work is submitted to SIGGRAPH 2021. Another rendering paper is also submitted to SIGGRAPH 2021 that develops a new theoretical model to improve Monte Carlo integration.