Vaibhav Vavilala
ABSTRACT
Delivering high resolution content is a challenge in the film and games industries due to the cost of photorealistic ray-traced rendering. Image upscaling techniques are commonly used to obtain a high resolution result from a low resolution render. Recently, deep learned upscaling has started to make an impact in production settings, synthesizing sharper and more detailed imagery than previous methods. The quality of a super resolution model depends on the size of its dataset, which can be expensive to generate at scale due to the large number of ray-traced pairs of renders required. In this report, we discuss our experiments training an additional neural network to learn the degradation operator, which can be used to rapidly generate low resolution images from existing high resolution renders. Our testing on production scenes shows that super resolution networks trained with a large synthetic dataset produce fewer artifacts and better reconstruction quality than networks trained on a smaller rendered dataset alone, and compare favorably to recent state of the art blind synthetic data techniques.
- Victor Cornillere, Abdelaziz Djelouah, Wang Yifan, Olga Sorkine-Hornung, and Christopher Schroers. 2019. Blind image super-resolution with spatially variant degradations. ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–13.Google Scholar
Digital Library
- Vaibhav Vavilala and Mark Meyer. 2020. Deep Learned Super Resolution for Feature Film Production. In ACM SIGGRAPH 2020 Talks(SIGGRAPH ’20). Association for Computing Machinery, New York, NY, USA, Article 41, 2 pages. https://doi.org/10.1145/3388767.3407334Google Scholar
- Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. 2018. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 864–873.Google ScholarCross Ref
Index Terms
Towards Large-Scale Super Resolution Datasets via Learned Downsampling of Ray-Traced Renderings
Recommendations
Image super-resolution: use of self-learning and gabor prior
ACCV'12: Proceedings of the 11th Asian conference on Computer Vision - Volume Part IIIRecent approaches on single image super-resolution (SR) have attempted to exploit self-similarity to avoid the use of multiple images. In this paper, we propose an SR method based on self-learning and Gabor prior. Given a low resolution (LR) test image ...
Single image super-resolution via self-similarity and low-rank matrix recovery
We propose a novel single-image super resolution (SISR) approach using self-similarity of image and the low-rank matrix recovery (LRMR). The method performs multiple upsampling steps with relatively small magnification factors to recover a desired high ...
Hyperspectral image super-resolution via convolutional neural network
2017 IEEE International Conference on Image Processing (ICIP)Due to the tradeoff between spatial and spectral resolution in remote sensing imaging, hyperspectral images are often acquired with a relative low spatial resolution, which limits their applications in many areas. Inspired by recent achievements in ...
Comments