RESEARCH TOPIC
Deep learning for photorealistic rendering
We have actively studied regression and learning techniques that estimate the ground truth image from sparse (noisy) samples to generate photo-realistic rendering images more efficiently. We have recently developed novel local regression-based techniques and deep-learning methods that significantly improve the quality of rendering images generated by Monte Carlo ray-tracing algorithms (e.g., path tracing).
Differentiable rendering and neural rendering
We have recently been conducting research on various inverse rendering problems. This research topic includes differentiable rendering techniques for optimizing scene parameters from rendered images, as well as neural rendering methods like the neural radiance field for synthesizing novel view images from sparse input images.
Effective sampling for ray tracing
Ray tracing is a fundamental technology widely adopted for photorealistic light transport algorithms. Nevertheless, achieving a converged and clean image through ray tracing requires a significant amount of computational time. We aim at accelerating ray tracing by developing novel sampling algorithms and optimization techniques.
Photorealistic integration of virtual objects into real scenes for AR
Visually integrating virtual objects into real environments is a crucial ingredient for photo-realistic augmented reality (AR). We have worked on novel techniques that capture the real scene's illumination and interactively render virtual objects using the estimated light to accomplish the research goal.