研究目的
To efficiently compute the Light Field by combining DIBR (Depth-Image-Based-Rendering) and traditional ray-tracing in an adaptive fashion to synthesize images, aiming to support real-time global illumination rendering for applications such as Virtual Reality (VR) and Augmented Reality (AR).
研究成果
The proposed adaptable Light Field generation process significantly speeds up Synthetic Light Field generation by exploiting shareable pixels among nearby images, reducing the number of samples to be traced during error correction. The approach is 2 to 3 times faster than traditional DIBR-based techniques with the same image quality, offering up to 3.24X speedup in simple scenes and about 2X speedup in complex scenes.
研究不足
The approach may not work well when the total number of samples is low, as noise tends to show up everywhere, which is inevitable when using any Monte Carlo based algorithm in low sample count. Traditional filtering approaches are applied to mitigate this issue.
1：Experimental Design and Method Selection:
The approach combines DIBR and traditional ray-tracing adaptively to synthesize images, measuring color errors during runtime to determine the balance between DIBR and Ray Tracing. A multi-level design is added to exploit shareable pixels among images for error removal control.
2：Sample Selection and Data Sources:
The Light Field consists of several key images called ray database, precomputed and stored in disk or memory. The display extracts colors from the ray database during runtime.
3：List of Experimental Equipment and Materials:
A PC with Intel i7-7700K CPU and 32GB of RAM, using a CPU renderer based on RadeonRays with OpenCL back-end.
4：Experimental Procedures and Operational Workflow:
The Light Field rendering process is divided into many levels with different numbers of samples and pixel positions. LevelTasks are generated with their own sharing level and total sample count. The LF Manager controls the rendering process, iterating several times before the Light Field images are ready.
5：Data Analysis Methods:
The error measurement is PSNR, generated by comparing the target image with the ground truth. Performance is compared between standard rendering and the multi-level approach under the same time budget.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
