Presentation

Scalable Neural Indoor Scene Rendering
SessionTechnical Papers
Event Type
Technical Paper
Research & Education
Virtual
Full Conference Supporter
Full Conference
Virtual Conference Supporter
Virtual Conference
Exhibitor Additional Full Conference
Exhibitor Full Conference
Time
Location
DescriptionWe propose an MLP-based scalable neural scene representation for large-scale indoor scenes with reflections. Its kernel is tile-specific, two-layer MLPs that can encode view-independent and view-dependent appearances separately, which can be trained distributively through a background sampling technique. Our method can achieve temporally coherent, high-frequency reflection rendering results in real time.