Real-time video analytics on high-resolution cameras has become a popular technology for various intelligent services like traffic control and crowd monitoring. While extensive work has been done on improving analytics accuracy with timing guarantees, virtually all of them target static viewpoint cameras. In this paper, we present Uirapuru, a novel framework for real-time, edge-based video analytics on high-resolution steerable cameras. The actuation performed by those cameras brings significant dynamism to the scene, presenting a critical challenge to existing popular approaches such as frame tiling. To address this problem, Uirapuru incorporates a comprehensive understanding of camera actuation into the system design paired with fast adaptive tiling at a per-frame level. We evaluate Uirapuru on a high-resolution video dataset, augmented by pan-tilt-zoom (PTZ) movements typical for steerable cameras and on real-world videos collected from an actual PTZ camera. Our experimental results show that Uirapuru provides up to 1.45× improvement in accuracy while respecting specified latency budgets or reaches up to 4.53× inference speedup with on-par accuracy compared to state-of-the-art static camera approaches.

Abstract While high accuracy is of paramount importance for deep learning (DL) inference, serving inference requests on time is equally critical but has not been carefully studied especially when the request has to be served over a dynamic wireless network at the edge. In this paper, we propose Jellyfish—a novel edge DL inference serving system that achieves soft guarantees for end-to-end inference latency service-level objectives (SLO). Jellyfish handles the network variability by utilizing both data and deep neural network (DNN) adaptation to conduct tradeoffs between accuracy and latency. Jellyfish features a new design that enables collective adaptation policies where the decisions for data and DNN adaptations are aligned and coordinated among multiple users with varying network conditions. We propose efficient algorithms to continuously map users and adapt DNNs at runtime, so that we fulfill latency SLOs while maximizing the overall inference accuracy. We further investigate dynamic DNNs, i.e., DNNs that encompass multiple architecture variants, and demonstrate their potential benefit through preliminary experiments. Our experiments based on a prototype implementation and real-world WiFi and LTE network traces show that Jellyfish can meet latency SLOs at around the 99th percentile while maintaining high accuracy.

While high accuracy is of paramount importance for deep learning (DL) inference, serving inference requests on time is equally critical but has not been carefully studied especially when the request has to be served over a dynamic wireless network at the edge. In this paper, we propose Jellyfish—a novel edge DL inference serving system that achieves soft guarantees on end-to-end inference latency often specified as a service-level objective (SLO). To handle the network variability, Jellyfish exploits both data and deep neural network (DNN) adaptation to conduct tradeoffs between accuracy and latency. Jellyfish features a new design that enables collective adaptation policies where the decisions for data and DNN adaptations are aligned and coordinated among multiple users with varying network conditions. We propose efficient algorithms to dynamically adapt DNNs and map users, so that we fulfill latency SLOs while maximizing the overall inference accuracy. Our experiments based on a prototype implementation and real-world WiFi and LTE network traces show that Jellyfish can meet latency SLOs at around the 99th percentile while maintaining high accuracy.

Abstract We present a scalable solution to render complex scenes from a large amount of viewpoints. While previous approaches rely either on a scene or a view hierarchy to process multiple elements together, we make full use of both, enabling sublinear performance in terms of views and scene complexity. By concurrently traversing the hierarchies, we efficiently find shared information among views to amortize rendering costs. One example application is many‐light global illumination. Our solution accelerates shadow map generation for virtual point lights, whose number can now be raised to over a million while maintaining interactive rates.

Abstract Stereoscopic 3D technology gives visual content creators a new dimension of design when creating images and movies. While useful for conveying emotion, laying emphasis on certain parts of the scene, or guiding the viewer's attention, editing stereo content is a challenging task. Not respecting comfort zones or adding incorrect depth cues, for example depth inversion, leads to a poor viewing experience. In this paper, we present a solution for editing stereoscopic content that allows an artist to impose disparity constraints and removes resulting depth conflicts using an optimization scheme. Using our approach, an artist only needs to focus on important high‐level indications that are automatically made consistent with the entire scene while avoiding contradictory depth cues and respecting viewer comfort.

Abstract Spectral Monte‐Carlo methods are currently the most powerful techniques for simulating light transport with wavelength‐dependent phenomena (e.g., dispersion, colored particle scattering, or diffraction gratings). Compared to trichromatic rendering, sampling the spectral domain requires significantly more samples for noise‐free images. Inspired by gradient‐domain rendering, which estimates image gradients, we propose spectral gradient sampling to estimate the gradients of the spectral distribution inside a pixel. These gradients can be sampled with a significantly lower variance by carefully correlating the path samples of a pixel in the spectral domain, and we introduce a mapping function that shifts paths with wavelength‐dependent interactions. We compute the result of each pixel by integrating the estimated gradients over the spectral domain using a one‐dimensional screened Poisson reconstruction. Our method improves convergence and reduces chromatic noise from spectral sampling, as demonstrated by our implementation within a conventional path tracer.

We propose a novel framework for photometric stereo (PS) under low-light conditions using uncalibrated near-light illumination. It operates on free-form video sequences captured with a minimalistic and affordable setup. We address issues such as albedo variations, shadowing, perspective projections, and camera noise. Our method uses specular spheres detected with a perspective-correcting Hough transform to robustly triangulate light positions in the presence of outliers via a least-squares approach. Furthermore, we propose an iterative reweighting scheme in combination with an ℓp-norm minimizer to robustly solve the calibrated near-light PS problem. In contrast to other approaches, our framework reconstructs depth, albedo (relative to light source intensity), and normals simultaneously and is demonstrated on synthetic and real-world scenes.

Monte-Carlo rendering algorithms have traditionally a high computational cost, because they rely on tracing up to billions of light paths through a scene to physically simulate light transport. Traditional path reusing amortizes the cost of path sampling over multiple pixels, but introduces visually unpleasant correlation artifacts and cannot handle scenes with specular light transport. We present gradient-domain path reusing , a novel unbiased Monte-Carlo rendering technique, which merges the concept of path reusing with the recently introduced idea of gradient-domain rendering. Since correlation is a key element in gradient sampling, it is a natural fit to be performed together with path reusing and we show that the typical artifacts of path reusing are significantly reduced by exploiting the gradient domain. Further, by employing the tools for shifting paths that were designed in the context of gradient-domain rendering over the last years, we can generalize path reusing to support arbitrary scenes including specular light transport. Our method is unbiased and currently the fastest converging unidirectional rendering technique outperforming conventional and gradient-domain path tracing by up to almost an order of magnitude.

Abstract Voxel‐based approaches are today's standard to encode volume data. Recently, directed acyclic graphs (DAGs) were successfully used for compressing sparse voxel scenes as well, but they are restricted to a single bit of (geometry) information per voxel. We present a method to compressarbitrarydata, such as colors, normals, or reflectance information. By decoupling geometry and voxel data via a novel mapping scheme, we are able to apply the DAG principle to encode the topology, while using a palette‐based compression for the voxel attributes, leading to a drastic memory reduction. Our method outperforms existing state‐of‐the‐art techniques and is well‐suited for GPU architectures. We achieve real‐time performance on commodity hardware for colored scenes with up to 17 hierarchical levels (a 128K3voxel resolution), which are stored fully in core.

Abstract The quality of shadow mapping is traditionally limited by texture resolution. We present a novel lossless compression scheme for high‐resolution shadow maps based on precomputed multiresolution hierarchies. Traditional multiresolution trees can compactly represent homogeneous regions of shadow maps at coarser levels, but require many nodes for fine details. By conservatively adapting the depth map, we can significantly reduce the tree complexity. Our proposed method offers high compression rates, avoids quantization errors, exploits coherency along all data dimensions, and is well‐suited for GPU architectures. Our approach can be applied for coherent shadow maps as well, enabling several applications, including high‐quality soft shadows and dynamic lights moving on fixed‐trajectories.

Abstract Multiresolution Hierarchies (MH) and Directed Acyclic Graphs (DAG) are two recent approaches for the compression of high‐resolution shadow information. In this paper, we introduce Merged Multiresolution Hierarchies (MMH), a novel data structure that unifies both concepts. An MMH leverages both hierarchical homogeneity exploited in MHs, as well as topological similarities exploited in DAG representations. We propose an efficient hash‐based technique to quickly identify and remove redundant subtree instances in a modified relative MH representation. Our solution remains lossless and significantly improves the compression rate compared to both preceding shadow map compression algorithms, while retaining the full run‐time performance of traditional MH representations.

Abstract Adaptive filtering techniques have proven successful in handling non‐uniform noise in Monte‐Carlo rendering approaches. A recent trend is to choose an optimal filter per pixel from a selection of non spatially‐varying filters. Nonetheless, the best filter choice is difficult to predict in the absence of a reference rendering. Our approach relies on the observation that the reconstruction error is locally smooth for a given filter. Hence, we propose to construct a dense error prediction from a small set of sparse but robust estimates. The filter selection is then formulated as a non‐local optimization problem, which we solve via graph cuts, to avoid visual artifacts due to inconsistent filter choices. Our approach does not impose any restrictions on the used filters, outperforms previous state‐of‐the‐art techniques and provides an extensible framework for future reconstruction techniques.

Abstract We present a fast reconstruction filtering method for images generated with Monte Carlo–based rendering techniques. Our approach specializes in reducing global illumination noise in the presence of depth‐of‐field effects at very low sampling rates and interactive frame rates. We employ edge‐aware filtering in the sample space to locally improve outgoing radiance of each sample. The improved samples are then distributed in the image plane using a fast, linear manifold‐based approach supporting very large circles of confusion. We evaluate our filter by applying it to several images containing noise caused by Monte Carlo–simulated global illumination, area light sources and depth of field. We show that our filter can efficiently denoise such images at interactive frame rates on current GPUs and with as few as 4–16 samples per pixel. Our method operates only on the colour and geometric sample information output of the initial rendering process. It does not make any assumptions on the underlying rendering technique and sampling strategy and can therefore be implemented completely as a post‐process filter.

In this paper, we present and investigate an approach to fast multilabel color image segmentation using convex optimization techniques. The presented model is in some ways related to the well-known Mumford-Shah model, but deviates in certain important aspects. The optimization problem has been designed with two goals in mind. The objective function should represent fundamental concepts of image segmentation, such as incorporation of weighted curve length and variation of intensity in the segmented regions, while allowing transformation into a convex concave saddle point problem that is computationally inexpensive to solve. This paper introduces such a model, the nontrivial transformation of this model into a convex-concave saddle point problem, and the numerical treatment of the problem. We evaluate our approach by applying our algorithm to various images and show that our results are competitive in terms of quality at unprecedentedly low computation times. Our algorithm allows high-quality segmentation of megapixel images in a few seconds and achieves interactive performance for low resolution images.

We propose the computation of a perceptual motion blur in videos. Our technique takes the predicted eye motion into account when watching the video. Compared to traditional motion blur recorded by a video camera our approach results in a perceptual blur that is closer to reality. This postprocess can also be used to simulate different shutter effects or for other artistic purposes. It handles real and artificial video input, is easy to compute and has a low additional cost for rendered content. We illustrate its advantages in a user study using eye tracking.

Reconstructing objects from monocular input video data is a challenging task. Since cues about shading and illumination are ambiguous from a single viewpoint, simultaneously distinguishing between material properties, texture, and illumination is infeasible. To properly reconstruct scene illumination a model of the object surface is mandatory. As most 3D surface reconstruction, surface refinement, and illumination reconstruction techniques rely on a diffuse object surface model, we propose a technique to robustly reconstruct surface albedo from monocular input video data and a 3D mesh representation of an object. For all frames of the video sequence we collect color samples for visible vertices under varying illumination directions and correct for possible ambient occlusion using mesh geometry information. We then use a two-staged clustering approach to recover the most plausible surface albedo for every mesh vertex. The presented approach does not require any a-priori illumination model and is evaluated using synthetic as well as real-world test sequences.

Abstract Light refracted by a dispersive interface leads to beautifully colored patterns that can be rendered faithfully with spectral Monte‐Carlo methods. Regrettably, results often suffer from chromatic noise or banding, requiring high sampling rates and large amounts of memory compared to renderers operating in some trichromatic color space. Addressing this issue, we introduce spectral ray differentials, which describe the change of light direction with respect to changes in the spectrum. In analogy with the classic ray and photon differentials, this information can be used for filtering in the spectral domain. Effectiveness of our approach is demonstrated by filtering for offline spectral light and path tracing as well as for an interactive GPU photon mapper based on splatting. Our results show considerably less chromatic noise and spatial aliasing while retaining good visual similarity to reference solutions with negligible overhead in the order of milliseconds.

Abstract We present a new data structure for object space partitioning that can be represented completely implicitly. The bounds of each node in the tree structure are recreated at run‐time from the scene objects contained therein. By applying a presorting procedure to the geometry, only a known fraction of the geometry is needed to locate the bounding planes of any node. We evaluate the impact of the implicit bounding plane representation and compare our algorithm to a classic bounding volume hierarchy. Though the representation is completely implicit, we still achieve interactive frame rates on commodity hardware.

Abstract Interactive computation of global illumination is a major challenge in current computer graphics research. Global illumination heavily affects the visual quality of generated images. It is therefore a key attribute for the perception of photo‐realistic images. Path tracing is able to simulate the physical behaviour of light using Monte Carlo techniques. However, the computational burden of this technique prohibits interactive rendering times on standard commodity hardware in high‐quality. Trying to solve the Monte Carlo integration with fewer samples results in characteristic noisy images. Global illumination filtering methods take advantage of the fact that the integral for neighbouring pixels may be very similar. Averaging samples of similar characteristics in screen‐space may approximate the correct integral, but may result in visible outliers. In this paper, we present a novel path tracing pipeline based on an edge‐aware filtering method for the indirect illumination which produces visually more pleasing results without noticeable outliers. The key idea is not to filter the noisy path traced images but to use it as a guidance to filter a second image composed from characteristic scene attributes that do not contain noise by default. We show that our approach better approximates the Monte Carlo integral compared to previous methods. Since the computation is carried out completely in screen‐space it is therefore applicable to fully dynamic scenes, arbitrary lighting and allows for high‐quality path tracing at interactive frame rates on commodity hardware.