TL;DR
This paper tackles the challenge of rendering multiple stereoscopic views in real-time for large-scale, glasses-free multiview parallax displays. They introduce a distributed rendering system using commodity PCs and graphics accelerators, exploring both optical and software multiplexing techniques to achieve scalable performance.
What Problem Does This Paper Address?
The dream of truly immersive 3D experiences, particularly glasses-free autostereoscopic displays, hinges on the ability to generate and render numerous perspective views of a scene simultaneously. Multiview parallax displays, which use barriers or lenticular sheets to direct different views to different observers, offer this glasses-free experience. However, generating the high-resolution data required for these displays, especially for large screens and many views (e.g., 16 views), demands a powerful and scalable rendering subsystem. Traditional rendering approaches struggle to keep up with the computational demands, leading to performance bottlenecks. This paper addresses this by proposing and evaluating a distributed rendering architecture designed to drive these complex displays efficiently.
Methods
The core of the paper lies in its exploration of distributed rendering strategies for multiview parallax displays. They leverage the Chromium stream-processing framework, which allows for non-invasive modification of OpenGL commands. Two primary multiplexing approaches are investigated:
- Optical Multiplexing: In this method, each view is rendered as a complete image and then optically combined by the display hardware (lenticular sheets). The distributed rendering system assigns one server PC to each view, with a "tilesort" SPU distributing rendering commands. This approach is straightforward but requires precise physical alignment and can be limited by projector placement.
- Software Multiplexing: Here, the rendering software itself interleaves sub-pixels to create the multiview effect. This is implemented using a tile-based architecture where each server PC renders all views for a specific tile of the display. This approach offers more flexibility in display design and can potentially reduce crosstalk. They developed custom SPUs within Chromium:
- Multiview SPU: This SPU on the client side transforms a single-view application command stream into a stream that renders a multiview frame, packing commands and signaling view changes.
- Compositing SPU: On the server side, this SPU interprets special commands to manage depth masks for discarding pixels not belonging to the current view and to blend tiles for smooth transitions.
They also detail the construction of novel large-scale multiview parallax display prototypes, including a rear-projection parallax-barrier display, and a novel automatic camera-based calibration procedure for precise alignment.
Datasets
Not discussed in the paper. The evaluation uses a simple test scene with an average of less than one thousand triangles per frame.
Key Insights and Ideas
- Chromium's Flexibility: The paper highlights Chromium's power as a platform for developing custom rendering solutions for specialized hardware like multiview displays without requiring application source code modifications.
- Optical vs. Software Multiplexing Trade-offs: Optical multiplexing offers superior performance (higher frame rates) but is less flexible in terms of display implementation and requires more complex projector setups. Software multiplexing, while potentially having higher network bandwidth overhead, allows for more compact and adaptable display designs.
- Calibration is Crucial: Precise geometric alignment of projectors and display elements is critical for high-quality multiview displays, and the paper presents a novel automated solution.
- Bottlenecks Emerge: As the number of views increases, performance is limited by CPU utilization on the application node, network bandwidth, or server rendering performance, depending on the multiplexing strategy.
Key Results
- The developed distributed rendering system successfully drives multiview parallax displays with 16 views.
- Optical multiplexing demonstrated superior performance, achieving higher rendered multiview frames per second compared to software multiplexing, especially as the number of views increased.
- Software multiplexing showed a performance drop that was more directly tied to the number of views, with the application node CPU and network bandwidth becoming significant bottlenecks.
- The paper presents the construction and evaluation of a novel rear-projection parallax-barrier display.
Limitations and Future Work
- Network Bandwidth and CPU Load: The current software multiplexing implementation exhibits significant network bandwidth overhead and high CPU utilization on the application node, especially with more than four views.
- Future Work: The authors plan to alleviate these bottlenecks by:
- Implementing a custom "tilesort" SPU to move multiview rendering completely to the server nodes.
- Improving the quality of the multiview parallax-barrier display by addressing interperspective aliasing.
- Further optimizing the performance of their Chromium SPUs.
Summary updated