Robert Herzog

New measurements reveal weaknesses of image quality metrics in evaluating graphics artifacts

Reliable detection of global illumination and rendering artifacts in the form of localized distortion maps is important for many graphics applications. Although many quality metrics have been developed for this task, they are often tuned for compression/transmission artifacts and have not been evaluated in the context of synthetic CG-images. In this work, we run two experiments where observers use a brush-painting interface to directly mark image regions with noticeable/objectionable distortions in the presence/absence of a high-quality reference image, respectively. The collected data shows a relatively high correlation between the with-reference and no-reference observer markings. Also, our demanding per-pixel image-quality datasets reveal weaknesses of both simple (PSNR, MSE, sCIE-Lab) and advanced (SSIM, MS-SSIM, HDR-VDP-2) quality metrics. The most problematic are excessive sensitivity to brightness and contrast changes, the calibration for near visibility-threshold distortions, lack of discrimination between plausible/implausible illumination, and poor spatial localization of distortions for multi-scale metrics. We believe that our datasets have further potential in improving existing quality metrics, but also in analyzing the saliency of rendering distortions, and investigating visual equivalence given our with- and no-reference data.

Spatio-temporal upsampling on the GPU

Pixel processing is becoming increasingly expensive for real-time applications due to the complexity of todays shaders and high-resolution framebuffers. However, most shading results are spatially or temporally coherent, which allows for sparse sampling and reuse of neighboring pixel values. This paper proposes a simple framework for spatio-temporal upsampling on modern GPUs. In contrast to previous work, which focuses either on temporal or spatial processing on the GPU, we exploit coherence in both. Our algorithm combines adaptive motion-compensated filtering over time and geometry-aware upsampling in image space. It is robust with respect to high-frequency temporal changes, and achieves substantial performance improvements by limiting the number of recomputed samples per frame. At the same time, we increase the quality of spatial upsampling by recovering missing information from previous frames. This temporal strategy also allows us to ensure that the image converges to a higher quality result.

NoRM: No-Reference Image Quality Metric for Realistic Image Synthesis

Synthetically generating images and video frames of complex 3D scenes using some photo‐realistic rendering software is often prone to artifacts and requires expert knowledge to tune the parameters. The manual work required for detecting and preventing artifacts can be automated through objective quality evaluation of synthetic images. Most practical objective quality assessment methods of natural images rely on a ground‐truth reference, which is often not available in rendering applications. While general purpose no‐reference image quality assessment is a difficult problem, we show in a subjective study that the performance of a dedicated no‐reference metric as presented in this paper can match the state‐of‐the‐art metrics that do require a reference. This level of predictive power is achieved exploiting information about the underlying synthetic scene (e.g., 3D surfaces, textures) instead of merely considering color, and training our learning framework with typical rendering artifacts. We show that our method successfully detects various non‐trivial types of artifacts such as noise and clamping bias due to insufficient virtual point light sources, and shadow map discretization artifacts. We also briefly discuss an inpainting method for automatic correction of detected artifacts.

Scalable Remote Rendering with Depth and Motion-flow Augmented Streaming

In this paper, we focus on efficient compression and streaming of frames rendered from a dynamic 3D model. Remote rendering and on‐the‐fly streaming become increasingly attractive for interactive applications. Data is kept confidential and only images are sent to the client. Even if the clients hardware resources are modest, the user can interact with state‐of‐the‐art rendering applications executed on the server. Our solution focuses on augmented video information, e.g., by depth, which is key to increase robustness with respect to data loss, image reconstruction, and is an important feature for stereo vision and other client‐side applications. Two major challenges arise in such a setup. First, the server workload has to be controlled to support many clients, second the data transfer needs to be efficient. Consequently, our contributions are twofold. First, we reduce the server‐based computations by making use of sparse sampling and temporal consistency to avoid expensive pixel evaluations. Second, our data‐transfer solution takes limited bandwidths into account, is robust to information loss, and compression and decompression are efficient enough to support real‐time interaction. Our key insight is to tailor our method explicitly for rendered 3D content and shift some computations on client GPUs, to better balance the server/client workload. Our framework is progressive, scalable, and allows us to stream augmented high‐resolution (e.g., HD‐ready) frames with small bandwidth on standard hardware.

Learning to Predict Localized Distortions in Rendered Images

In this work, we present an analysis of feature descriptors for objective image quality assessment. We explore a large space of possible features including components of existing image quality metrics as well as many traditional computer vision and statistical features. Additionally, we propose new features motivated by human perception and we analyze visual saliency maps acquired using an eye tracker in our user experiments. The discriminative power of the features is assessed by means of a machine learning framework revealing the importance of each feature for image quality assessment task. Furthermore, we propose a new data-driven full-reference image quality metric which outperforms current state-of-theart metrics. The metric was trained on subjective ground truth data combining two publicly available datasets. For the sake of completeness we create a new testing synthetic dataset including experimentally measured subjective distortion maps. Finally, using the same machine-learning framework we optimize the parameters of popular existing metrics.

Fast Final Gathering via Reverse Photon Mapping

We present a new algorithm for computing indirect illumination based on density estimation similarly to photon mapping. We accelerate the search for nal gathering by reorganizing the computation in the reverse order. We use two trees that organize spatially not only the position of photons but also the position of nal gather rays. The achieved speedup is algorithmic, the performance improvement takes advantage of logarithmic complexity of searching in trees. The algorithm requires almost no user settings unlike many known acceleration techniques for photon mapping. The image quality is the same as for traditional photon mapping with nal gathering, since the algorithm does not approximate or interpolate. Optionally, the algorithm can be combined with other techniques such as density control and importance sampling. The algorithm creates a coherent access pattern to the main memory. This further improves on performance and also allows us to use efcient external data structures to alleviate the increased memory requirements.

Ray maps for global illumination

We describe a novel data structure for representing light transport called ray map. The ray map extends the concept of photon maps: it stores not only photon impacts but the whole photon paths. We demonstrate the utility of ray maps for global illumination by eliminating boundary bias and reducing topological bias of density estimation in global illumination. Thanks to the elimination of boundary bias we could use ray maps for fast direct visualization with the image quality being close to that obtained by the expensive final gathering step. We describe in detail our implementation of the ray map using a lazily constructed kD-tree. We also present several optimizations bringing the ray map query performance close to the performance of the photon map.

Perceptual depth compression for stereo applications

Conventional depth video compression uses video codecs designed for color images. Given the performance of current encoding standards, this solution seems efficient. However, such an approach suffers from many issues stemming from discrepancies between depth and light perception. To exploit the inherent limitations of human depth perception, we propose a novel depth compression method that employs a disparity perception model. In contrast to previous methods, we account for disparity masking, and model a distinct relation between depth perception and contrast in luminance. Our solution is a natural extension to the H.264 codec and can easily be integrated into existing decoders. It significantly improves both the compression efficiency without sacrificing visual quality of depth of rendered content, and the output of depth‐reconstruction algorithms or depth cameras.

Anisotropic Radiance-Cache Splatting for Efficiently Computing High-Quality Global Illumination with Lightcuts

Computing global illumination in complex scenes is even with todays computational power a demanding task. In this work we propose a novel irradiance caching scheme that combines the advantages of two state‐of‐the‐art algorithms for high‐quality global illumination rendering: lightcuts, an adaptive and hierarchical instant‐radiosity based algorithm and the widely used (ir)radiance caching algorithm for sparse sampling and interpolation of (ir)radiance in object space. Our adaptive radiance caching algorithm is based on anisotropic cache splatting, which adapts the cache footprints not only to the magnitude of the illumination gradient computed with light‐cuts but also to its orientation allowing larger interpolation errors along the direction of coherent illumination while reducing the error along the illumination gradient. Since lightcuts computes the direct and indirect lighting seamlessly, we use a two‐layer radiance cache, to store and control the interpolation of direct and indirect lighting individually with different error criteria. In multiple iterations our method detects cache interpolation errors above the visibility threshold of a pixel and reduces the anisotropic cache footprints accordingly. We achieve significantly better image quality while also speeding up the computation costs by one to two orders of magnitude with respect to the well‐known photon mapping with (ir)radiance caching procedure.

Render2MPEG: A Perception‐based Framework Towards Integrating Rendering and Video Compression

Currently 3D animation rendering and video compression are completely independent processes even if rendered frames are streamed on‐the‐fly within a client‐server platform. In such scenario, which may involve time‐varying transmission bandwidths and different display characteristics at the client side, dynamic adjustment of the rendering quality to such requirements can lead to a better use of server resources. In this work, we present a framework where the renderer and MPEG codec are coupled through a straightforward interface that provides precise motion vectors from the rendering side to the codec and perceptual error thresholds for each pixel in the opposite direction. The perceptual error thresholds take into account bandwidth‐dependent quantization errors resulting from the lossy com‐pression as well as image content‐dependent luminance and spatial contrast masking. The availability of the discrete cosine transform (DCT) coefficients at the codec side enables to use advanced models of the human visual system (HVS) in the perceptual error threshold derivation without incurring any significant cost. Those error thresholds are then used to control the rendering quality and make it well aligned with the compressed stream quality. In our prototype system we use the lightcuts technique developed by Walter et al., which we enhance to handle dynamic image sequences, and an MPEG‐2 implementation. Our results clearly demonstrate many advantages of coupling the rendering with video compression in terms of faster rendering. Furthermore, temporally coherent rendering leads to a reduction of temporal artifacts.