IDLab-MEDIA
Dimensionality Reduction for Real-Time Light-Field View Synthesis of Kernel-Based Models

Dimensionality Reduction for Real-Time Light-Field View Synthesis of Kernel-Based Models

We introduce a novel gpu-accelerated real-time 4D planar light-field renderer based on Steered Mixture-of-Experts (SMoE). Our proposed GPU graphics pipeline achieves real-time view-synthesis while retaining SMoE’s quality almost perfectly. At 180 to 290 frames per second for a resolution of 2048x2048 pixels (using an NVIDIA RTX 2080Ti), our method is VR-ready and by far the fastest kernel-based renderer.

Results video of spin movement within Barbershop scene.

Additionally our proposed method:

  • has a minimal memory and disk footprint with scenes typically being under 5 MB.
  • has rendering performance independent of the scene’s geometry, making performance consistent across scenes.
  • is the only kernel based method with linear time complexity in both the number of pixels and the number of components.

Our results can be downloaded here.