Motion Blur for EWA Surface Splatting

Simon Heinzle, Johanna Wolf, Yoshihiro Kanamori, Tim Weyrich, Tomoyuki Nishita, Markus H. GrossComputer Graphics Forum 2010 · DOI

TL;DR

This paper introduces a novel framework for rendering motion blur in point-based graphics using Elliptical Weighted Average (EWA) surface splatting. By extending EWA splats into the temporal dimension with 3D kernels, they achieve high-quality, realistic motion blur for dynamic scenes at interactive frame rates, overcoming limitations of previous supersampling techniques.

What Problem Does This Paper Address?

Rendering realistic motion blur in computer graphics is crucial for conveying dynamism and realism in animations and simulations. Traditional methods often rely on temporal supersampling, which involves rendering many frames and averaging them, leading to significant computational cost. For point-based rendering techniques like EWA surface splatting, which excel at rendering complex geometry with high visual quality, incorporating motion blur without sacrificing performance or introducing artifacts has been a challenge. Existing methods either struggle with complex motion, inter-object occlusions, or require computationally expensive supersampling. This paper aims to bridge this gap by developing a principled approach to motion blur within the EWA framework that is both efficient and visually accurate.

Methods

The core innovation of this paper is the extension of the 2D EWA surface splatting framework into the temporal domain. Instead of using 2D Gaussian kernels to reconstruct surfaces at a single point in time, the authors introduce 3D spatio-temporal reconstruction kernels. These kernels are essentially ellipsoids that encompass both the spatial extent of a surface point and its motion trajectory over a short time interval.

Key methodological contributions include:

  • 3D Spatio-Temporal Kernels: The 2D EWA kernels are extended into 3D by convolving them with a 1D Gaussian along the instantaneous motion vector. This creates ellipsoidal primitives that naturally capture the motion of surface points.
  • Mathematical Framework: The paper derives a new mathematical framework for motion-blurred EWA surface splatting, unifying spatial and temporal components. This involves defining a continuous spatio-temporal screen-space signal and bandlimiting it with a spatio-temporal anti-aliasing filter.
  • Rendering Algorithm: A rendering algorithm is proposed that leverages these 3D kernels. It involves projecting these ellipsoidal kernels to screen space and accumulating their contributions.
  • Visibility Approximation: A crucial aspect is handling visibility in dynamic scenes. The authors propose an A-Buffer-like approach, implemented in multiple GPU passes, to approximate the temporal visibility of reconstruction kernels. This involves determining the earliest and latest times a kernel is visible for each pixel.
  • GPU Implementation: The algorithm is implemented using a six-pass GPU pipeline, utilizing vertex, geometry, and fragment shaders to achieve interactive frame rates. This includes specialized passes for depth buffer computation, background visibility estimation, kernel blending, and normalization.

Datasets

Not discussed in the paper. The paper focuses on a novel rendering technique and evaluates it using synthetic scenes generated by the authors, showcasing various dynamic objects like dragons, knots, faces, and books.

Key Insights and Ideas

  • Unifying Space and Time: The most significant insight is the elegant unification of spatial and temporal components into a single 3D kernel. This allows for a more principled and integrated approach to motion blur in point-based rendering.
  • Ellipsoids as Motion Primitives: Using ellipsoids as rendering primitives that inherently capture motion is a powerful concept. It allows the rendering system to adapt to varying motion speeds without explicit temporal supersampling of the entire scene.
  • GPU-Accelerated Visibility Approximation: The multi-pass GPU approach for approximating temporal visibility is a key enabler for interactive performance. While not perfectly accurate, it strikes a good balance between visual quality and speed.
  • Trade-off Between Speed and Quality: The paper highlights that increasing the temporal sampling of motion paths (similar to supersampling) can improve accuracy at the cost of performance, offering a flexible trade-off.

Key Results

The paper demonstrates the ability to produce high-quality, realistic motion blur for point-based geometry at interactive frame rates.

  • Visual Quality: The rendered images (Figures 1, 7, 8, 9) showcase smooth and accurate motion blur, effectively capturing the dynamics of moving objects. The results are visually superior to traditional temporal supersampling in terms of smoothness and artifact reduction.
  • Performance: The GPU implementation achieves competitive frame rates, often outperforming temporal supersampling for comparable visual quality (Table 1). For example, a scene with 350k points rendered with 3D kernels at 1.80M points/s is compared to supersampling at 1.44M points/s with a factor of 30x.
  • Handling Varying Motion: The volumetric kernels adapt to varying motion speeds, which is a significant advantage over fixed-rate supersampling.

Limitations and Future Work

  • Inter-Object Occlusions: The primary limitation acknowledged is the difficulty in accurately handling inter-object occlusions on the GPU due to the approximate visibility computation. This can lead to artifacts, especially when objects overlap in both time and depth.
  • Approximated Visibility: The visibility approximation, which relies on start and end frame depth information, can lead to artifacts when geometry is only visible for short durations within the exposure interval.
  • Future Work:
    • More Accurate Visibility: The authors suggest exploring more advanced GPU programming techniques (e.g., compute shaders) or specialized hardware extensions for more accurate visibility computation.
  • Adaptive Sampling: Varying the temporal sampling rate adaptively for each point based on its motion and shading changes could improve quality.
  • Improved Occlusion Handling: Developing better methods for handling inter-object relationships is crucial for more complex scenes.

Summary updated

BibTeX
@article{heinzle2010motion,
  author = {Simon Heinzle and Johanna Wolf and Yoshihiro Kanamori and Tim Weyrich and Tomoyuki Nishita and Markus H. Gross},
  title = {{Motion Blur for EWA Surface Splatting}},
  journal = {Computer Graphics Forum},
  year = {2010},
  doi = {10.1111/j.1467-8659.2009.01643.x},
}

Loading related content...

Figure