Parametric functions are an extremely efficient representation for 3D geometry, capable of compactly modelling highly complex objects. Once specified, parametric 3D objects allow for visualization at arbitrary levels of detail (LOD), at no additional memory cost, limited only by the amount of evaluated samples. However, mapping the sample evaluation to the hardware rendering pipelines of modern graphics processing units (GPUs) is not trivial. In this article, we propose a general method for efficient rendering of parametrically-defined 3D objects on modern hardware architectures. Our method adaptively analyzes, allocates and evaluates parametric function samples to produce high-quality renderings. Geometric precision can be modulated from few pixels down to sub-pixel level, enabling real-time frame rates of several 100 frames per second (FPS) for various parametric functions. We propose a dedicated LOD stage, which outputs patches of similar geometric detail to a subsequent rendering stage that uses either a hardware tessellation-based approach or performs point-based software rasterization. Our method requires neither preprocessing nor caching, and the proposed LOD mechanism is fast enough to run each frame. Hence, our approach also lends itself to animated parametric objects. We demonstrate the benefits of our method over a state-of-the-art spherical harmonics (SH) glyph rendering method and over classical LOD approaches, while showing its flexibility on a range of other demanding shapes.

Decomposing geometry, materials and lighting from a set of images, namely inverse rendering, has been a long-standing problem in computer vision and graphics. Recent advances in neural rendering enable photo-realistic and plausible inverse rendering results. The emergence of 3D Gaussian Splatting has boosted it to the next level by showing real-time rendering potentials. An intuitive finding is that the models used for inverse rendering do not take into account the dependency of opacity w.r.t. material properties, namely cross section, as suggested by optics. Therefore, we develop a novel approach that adds this dependency to the modeling itself. Inspired by radiative transfer, we augment the opacity term by introducing a neural network that takes as input material properties to provide modeling of cross section and a physically correct activation function. The gradients for material properties are therefore not only from color but also from opacity, facilitating a constraint for their optimization. Therefore, the proposed method incorporates more accurate physical properties compared to previous works. We implement our method into 3 different baselines that use Gaussian Splatting for inverse rendering and achieve significant improvements universally in terms of novel view synthesis and material modeling.

Abstract Since its introduction, 3D Gaussian Splatting (3DGS) has become an important reference method for learning 3D representations of a captured scene, allowing real‐time novel‐view synthesis with high visual quality and fast training times. Neural Radiance Fields (NeRFs), which preceded 3DGS, are based on a principled ray‐marching approach for volumetric rendering. In contrast, while sharing a similar image formation model with NeRF, 3DGS uses a hybrid rendering solution that builds on the strengths of volume rendering and primitive rasterization. A crucial benefit of 3DGS is its performance, achieved through a set of approximations, in many cases with respect to volumetric rendering theory. A naturally arising question is whether replacing these approximations with more principled volumetric rendering solutions can improve the quality of 3DGS. In this paper, we present an in‐depth analysis of the various approximations and assumptions used by the original 3DGS solution. We demonstrate that, while more accurate volumetric rendering can help for low numbers of primitives, the power of efficient optimization and the large number of Gaussians allows 3DGS to outperform volumetric rendering despite its approximations.

Human-centered live events have always played a pivotal role in shaping culture and fostering social connections. Traditional 2D live transmissions fail to replicate the immersive quality of physical attendance. Addressing this gap, this paper proposes LiveSplats , a framework towards real-time, photo-realistic 3D reconstructions of live events using high-performance 3D Gaussian Splatting. Our solution capitalizes on strong geometric priors to optimize through distributed processing and load balancing, enabling interactive, freely explorable 3D experiences. By dividing scene reconstruction into actor-centric and environment-specific tasks, we employ hierarchical coarse-to-fine optimization to rapidly and accurately reconstruct human actors based on pose data, refining their geometry and appearance with photometric loss. For static environments, we focus on view-dependent appearance changes, streamlining rendering efficiency and maximizing GPU performance. To facilitate evaluation, we introduce (and distribute) a synthetic benchmark dataset of basketball games, offering high visual fidelity as ground truth. In both our synthetic benchmark and publicly available benchmarks, LiveSplats consistently outperforms existing approaches. The dataset is available at https://humansensinglab.github.io/basket-multiview.

Radiance field methods such as 3D Gaussian Splatting (3DGS) allow easy reconstruction from photos, enabling free-viewpoint navigation. Nonetheless, pose estimation using Structure from Motion and 3DGS optimization can still each take between minutes and hours of computation after capture is complete. SLAM methods combined with 3DGS are fast but struggle with wide camera baselines and large scenes. We present an on-the-fly method to produce camera poses and a trained 3DGS immediately after capture. Our method can handle dense and wide-baseline captures of ordered photo sequences and large-scale scenes. To do this, we first introduce fast initial pose estimation, exploiting learned features and a GPU-friendly mini bundle adjustment. We then introduce direct sampling of Gaussian primitive positions and shapes, incrementally spawning primitives where required, significantly accelerating training. These two efficient steps allow fast and robust joint optimization of poses and Gaussian primitives. Our incremental approach handles large-scale scenes by introducing scalable radiance field construction, progressively clustering 3DGS primitives, storing them in anchors, and offloading them from the GPU. Clustered primitives are progressively merged, keeping the required scale of 3DGS at any viewpoint. We evaluate our solution on a variety of datasets and show that it can provide on-the-fly processing of all the capture scenarios and scene sizes we target. At the same time our method remains competitive - in speed, image quality, or both - with other methods that only handle specific capture styles or scene sizes.

Abstract We propose MatUp, an upsampling filter for material super‐resolution. Our method takes as input a low‐resolution SVBRDF and upscales its maps so that their rendering under various lighting conditions fits upsampled renderings inferred in the radiance domain with pre‐trained RGB upsamplers. We formulate our local filter as a compact Multilayer Perceptron (MLP), which acts on a small window of the input SVBRDF and is optimized using a data‐fitting loss defined over upsampled radiance at various locations. This optimization is entirely performed at the scale of a single, independent material. Doing so, MatUpleverages the reconstruction capabilities acquired over large collections of natural images by pre‐trained RGB models and provides regularization over self‐similar structures. In particular, our light‐weight neural filter avoids retraining complex architectures from scratch or accessing any large collection of low/high resolution material pairs – which do not actually exist at the scale RGB upsamplers are trained with. As a result, MatUpprovides fine and coherent details in the upscaled material maps, as shown in the extensive evaluation we provide.

Abstract Terrain analysis plays an important role in computer graphics, hydrology and geomorphology. In particular, analyzing the path of material flow over a terrain with consideration of local depressions is a precursor to many further tasks in erosion, river formation, and plant ecosystem simulation. For example, fluvial erosion simulation used in terrain modeling computes water discharge to repeatedly locate erosion channels for soil removal and transport. Despite its significance, traditional methods face performance constraints, limiting their broader applicability. In this paper, we propose a novel GPU flow routing algorithm that computes the water discharge in 𝒪(log n) iterations for a terrain with n vertices (assuming n processors). We also provide a depression routing algorithm to route the water out of local minima formed by depressions in the terrain, which converges in 𝒪(log2 n) iterations. Our implementation of these algorithms leads to a 5× speedup for flow routing and 34 × to 52 × speedup for depression routing compared to previous work on a 10242 terrain, enabling interactive control of terrain simulation.

Abstract Radiance field methods represent the state of the art in reconstructing complex scenes from multi‐view photos. However, these reconstructions often suffer from one or both of the following limitations: First, they typically represent scenes in low dynamic range (LDR), which restricts their use to evenly lit environments and hinders immersive viewing experiences. Secondly, their reliance on a pinhole camera model, assuming all scene elements are in focus in the input images, presents practical challenges and complicates refocusing during novel‐view synthesis. Addressing these limitations, we present a lightweight method based on 3D Gaussian Splatting that utilizes multi‐view LDR images of a scene with varying exposure times, apertures, and focus distances as input to reconstruct a high‐dynamic‐range (HDR) radiance field. By incorporating analytical convolutions of Gaussians based on a thin‐lens camera model as well as a tonemapping module, our reconstructions enable the rendering of HDR content with flexible refocusing capabilities. We demonstrate that our combined treatment of HDR and depth of field facilitates real‐time cinematic rendering, outperforming the state of the art.

Novel view synthesis has seen major advances in recent years, with 3D Gaussian splatting offering an excellent level of visual quality, fast training and real-time rendering. However, the resources needed for training and rendering inevitably limit the size of the captured scenes that can be represented with good visual quality. We introduce a hierarchy of 3D Gaussians that preserves visual quality for very large scenes, while offering an efficient Level-of-Detail (LOD) solution for efficient rendering of distant content with effective level selection and smooth transitions between levels. We introduce a divide-and-conquer approach that allows us to train very large scenes in independent chunks. We consolidate the chunks into a hierarchy that can be optimized to further improve visual quality of Gaussians merged into intermediate nodes. Very large captures typically have sparse coverage of the scene, presenting many challenges to the original 3D Gaussian splatting training method; we adapt and regularize training to account for these issues. We present a complete solution, that enables real-time rendering of very large scenes and can adapt to available resources thanks to our LOD method. We show results for captured scenes with up to tens of thousands of images with a simple and affordable rig, covering trajectories of up to several kilometers and lasting up to one hour.

Gaussian Splatting has emerged as a prominent model for constructing 3D representations from images across diverse domains. However, the efficiency of the 3D Gaussian Splatting rendering pipeline relies on several simplifications. Notably, reducing Gaussian to 2D splats with a single viewspace depth introduces popping and blending artifacts during view rotation. Addressing this issue requires accurate per-pixel depth computation, yet a full per-pixel sort proves excessively costly compared to a global sort operation. In this paper, we present a novel hierarchical rasterization approach that systematically resorts and culls splats with minimal processing overhead. Our software rasterizer effectively eliminates popping artifacts and view inconsistencies, as demonstrated through both quantitative and qualitative measurements. Simultaneously, our method mitigates the potential for cheating view-dependent effects with popping, ensuring a more authentic representation. Despite the elimination of cheating, our approach achieves comparable quantitative results for test images, while increasing the consistency for novel view synthesis in motion. Due to its design, our hierarchical approach is only 4% slower on average than the original Gaussian Splatting. Notably, enforcing consistency enables a reduction in the number of Gaussians by approximately half with nearly identical quality and view-consistency. Consequently, rendering performance is nearly doubled, making our approach 1.6x faster than the original Gaussian Splatting, with a 50% reduction in memory requirements. Our renderer is publicly available at https://github.com/r4dl/StopThePop.

For over two decades, the OpenGL API provided users with the means for implementing versatile, feature-rich, and portable real-time graphics applications. Consequently, it has been widely adopted by practitioners and educators alike and is deeply ingrained in many curricula that teach real-time graphics for higher education. Over the years, the architecture of graphics processing units (GPUs) incrementally diverged from OpenGL’s conceptual design. The more recently introduced Vulkan API provides a more modern, fine-grained approach for interfacing with the GPU, which allows a high level of controllability and, thereby, deep insights into the inner workings of modern GPUs. This property makes the Vulkan API especially well suitable for teaching graphics programming in university education, where fundamental knowledge shall be conveyed. Hence, it stands to reason that educators who have their students’ best interests at heart should provide them with corresponding lecture material. However, Vulkan is notoriously verbose and rather challenging for first-time users, thus transitioning to this new API bears a considerable risk of failing to achieve expected teaching goals. In this paper, we document our experiences after teaching Vulkan in both introductory and advanced graphics courses side-by-side with conventional OpenGL. A collection of surveys enables us to draw conclusions about perceived workload, difficulty, and students’ acceptance of either approach. In doing so, we identify suitable conditions and recommendations for teaching Vulkan to both undergraduate and graduate students.

Abstract About: We introduce a GPU‐accelerated LOD construction process that creates a hybrid voxel‐point‐based variation of the widely used layered point cloud (LPC) structure for LOD rendering and streaming. The massive performance improvements provided by the GPU allow us to improve the quality of lower LODs via color filtering while still increasing construction speed compared to the non‐filtered, CPU‐based state of the art. Background: LOD structures are required to render hundreds of millions to trillions of points, but constructing them takes time. Results: LOD structures suitable for rendering and streaming are constructed at rates of about 1 billion points per second (with color filtering) to 4 billion points per second (sample‐picking/random sampling, state of the art) on an RTX 3090 – an improvement of a factor of 80 to 400 times over the CPU‐based state of the art (12 million points per second). Due to being in‐core, model sizes are limited to about 500 million points per 24GB memory. Discussion: Our method currently focuses on maximizing in‐core construction speed on the GPU. Issues such as out‐of‐core construction of arbitrarily large data sets are not addressed, but we expect it to be suitable as a component of bottom‐up out‐of‐core LOD construction schemes.

Radiance Field methods have recently revolutionized novel-view synthesis of scenes captured with multiple photos or videos. However, achieving high visual quality still requires neural networks that are costly to train and render, while recent faster methods inevitably trade off speed for quality. For unbounded and complete scenes (rather than isolated objects) and 1080p resolution rendering, no current method can achieve real-time display rates. We introduce three key elements that allow us to achieve state-of-the-art visual quality while maintaining competitive training times and importantly allow high-quality real-time (≥ 30 fps) novel-view synthesis at 1080p resolution. First, starting from sparse points produced during camera calibration, we represent the scene with 3D Gaussians that preserve desirable properties of continuous volumetric radiance fields for scene optimization while avoiding unnecessary computation in empty space; Second, we perform interleaved optimization/density control of the 3D Gaussians, notably optimizing anisotropic covariance to achieve an accurate representation of the scene; Third, we develop a fast visibility-aware rendering algorithm that supports anisotropic splatting and both accelerates training and allows realtime rendering. We demonstrate state-of-the-art visual quality and real-time rendering on several established datasets.

Visibility computation is a key element in computer graphics applications. More specifically, a from-region potentially visible set (PVS) is an established tool in rendering acceleration, but its high computational cost means a from-region PVS is almost always precomputed. Precomputation restricts the use of PVS to static scenes and leads to high storage cost, in particular, if we need fine-grained regions. For dynamic applications, such as streaming content over a variable-bandwidth network, online PVS computation with configurable region size is required. We address this need with trim regions, a new method for generating from-region PVS for arbitrary scenes in real time. Trim regions perform controlled erosion of object silhouettes in image space, implicitly applying the shrinking theorem known from previous work. Our algorithm is the first that applies automatic shrinking to unconstrained 3D scenes, including non-manifold meshes, and does so in real time using an efficient GPU execution model. We demonstrate that our algorithm generates a tight PVS for complex scenes and outperforms previous online methods for from-viewpoint and from-region PVS. It runs at 60 Hz for realistic game scenes consisting of millions of triangles and computes PVS with a tightness matching or surpassing existing approaches.

Streaming rendered 3D content over a network to a thin client device, such as a phone or a VR/AR headset, brings high-fidelity graphics to platforms where it would not normally possible due to thermal, power, or cost constraints. Streamed 3D content must be transmitted with a representation that is both robust to latency and potential network dropouts. Transmitting a video stream and reprojecting to correct for changing viewpoints fails in the presence of disocclusion events; streaming scene geometry and performing high-quality rendering on the client is not possible on limited-power mobile GPUs. To balance the competing goals of disocclusion robustness and minimal client workload, we introduce QuadStream , a new streaming content representation that reduces motion-to-photon latency by allowing clients to efficiently render novel views without artifacts caused by disocclusion events. Motivated by traditional macroblock approaches to video codec design, we decompose the scene seen from positions in a view cell into a series of quad proxies , or view-aligned quads from multiple views. By operating on a rasterized G-Buffer, our approach is independent of the representation used for the scene itself; the resulting QuadStream is an approximate geometric representation of the scene that can be reconstructed by a thin client to render both the current view and nearby adjacent views. Our technical contributions are an efficient parallel quad generation, merging, and packing strategy for proxy views covering potential client movement in a scene; a packing and encoding strategy that allows masked quads with depth information to be transmitted as a frame-coherent stream; and an efficient rendering approach for rendering our QuadStream representation into entirely novel views on thin clients. We show that our approach achieves superior quality compared both to video data streaming methods, and to geometry-based streaming.

Abstract In this paper, we present several compute‐based point cloud rendering approaches that outperform the hardware pipeline by up to an order of magnitude and achieve significantly better frame times than previous compute‐based methods. Beyond basic closest‐point rendering, we also introduce a fast, high‐quality variant to reduce aliasing. We present and evaluate several variants of our proposed methods with different flavors of optimization, in order to ensure their applicability and achieve optimal performance on a range of platforms and architectures with varying support for novel GPU hardware features. During our experiments, the observed peak performance was reached rendering 796 million points (12.7GB) at rates of 62 to 64 frames per second (50 billion points per second, 802GB/s) on an RTX 3090 without the use of level‐of‐detail structures. We further introduce an optimized vertex order for point clouds to boost the efficiency of GL_POINTS by a factor of 5× in cases where hardware rendering is compulsory. We compare different orderings and show that Morton sorted buffers are faster for some viewpoints, while shuffled vertex buffers are faster in others. In contrast, combining both approaches by first sorting according to Morton‐code and shuffling the resulting sequence in batches of 128 points leads to a vertex buffer layout with high rendering performance and low sensitivity to viewpoint changes.

Abstract Following recent advances in GPU hardware development and newly introduced rendering pipeline extensions, the segmentation of input geometry into small geometry clusters—so‐called meshlets—has emerged as an important practice for efficient rendering of complex 3D models. Meshlets can be processed efficiently using mesh shaders on modern graphics processing units, in order to achieve streamlined geometry processing in just two tightly coupled shader stages that allow for dynamic workload manipulation in‐between. The additional granularity layer between entire models and individual triangles enables new opportunities for fine‐grained visibility culling methods. However, in contrast to static models, view frustum and backface culling on a per‐meshlet basis for skinned, animated models are difficult to achieve while respecting the conservative spatio‐temporal bounds that are required for robust rendering results. In this paper, we describe a solution for computing and exploiting relevant conservative bounds for culling meshlets of models that are animated using linear blend skinning. By enabling visibility culling for animated meshlets, our approach can help to improve rendering performance and alleviate bottlenecks in the notoriously performance‐ and memory‐intensive skeletal animation pipelines of modern real‐time graphics applications.

With the introduction of hardware-supported ray tracing and deep learning for denoising, computer graphics has made a considerable step toward real-time global illumination. In this work, we present an alternative global illumination method: The stochastic substitute tree (SST), a hierarchical structure inspired by lightcuts with light probability distributions as inner nodes. Our approach distributes virtual point lights (VPLs) in every frame and efficiently constructs the SST over those lights by clustering according to Morton codes. Global illumination is approximated by sampling the SST and considers the BRDF at the hit location as well as the SST nodes’ intensities for importance sampling directly from inner nodes of the tree. To remove the introduced Monte Carlo noise, we use a recurrent autoencoder. In combination with temporal filtering, we deliver real-time global illumination for complex scenes with challenging light distributions.

Harnessing the power of massively parallel devices like the graphics processing unit (GPU) is difficult for algorithms that show dynamic or inhomogeneous workloads. To achieve high performance, such advanced algorithms require scalable, concurrent queues to collect and distribute work. We show that previous queuing approaches are unfit for this task, as they either (1) do not work well in a massively parallel environment, or (2) obstruct the use of individual threads on top of single-instruction-multiple-data (SIMD) cores, or (3) block during access, thus prohibiting multi-queue setups. With these issues in mind, we present the Broker Queue, a highly efficient, fully linearizable FIFO queue for fine-granular parallel work distribution on the GPU. We evaluate its performance and usability on modern GPU models against a wide range of existing algorithms. The Broker Queue is up to three orders of magnitude faster than nonblocking queues and can even outperform significantly simpler techniques that lack desired properties for fine-granular work distribution.

Harnessing the power of massively parallel devices like the graphics processing unit (GPU) is difficult for algorithms that show dynamic or inhomogeneous workloads. To achieve high performance, such advanced algorithms require scalable, concurrent queues to collect and distribute work. We present a new concurrent work queue, the Broker Queue, a highly efficient, linearizable queue for fine-granular work distribution on the GPU. We evaluate its usability and benefits in contrast to existing queuing algorithms. Our queue is up to one order of magnitude faster than non-blocking queues, and outperforms simpler queue designs that are unfit for fine-granular work distribution.

In this paper, we present a real-time graphics pipeline implemented entirely in software on a modern GPU. As opposed to previous work, our approach features a fully-concurrent, multi-stage, streaming design with dynamic load balancing, capable of operating efficiently within bounded memory. We address issues such as primitive order, vertex reuse, and screen-space derivatives of dependent variables, which are essential to real-world applications, but have largely been ignored by comparable work in the past. The power of a software approach lies in the ability to tailor the graphics pipeline to any given application. In exploration of this potential, we design and implement four novel pipeline modifications. Evaluation of the performance of our approach on more than 100 real-world scenes collected from video games shows rendering speeds within one order of magnitude of the hardware graphics pipeline as well as significant improvements over previous work, not only in terms of capabilities and performance, but also robustness.

Abstract While the modern graphics processing unit (GPU) offers massive parallel compute power, the ability to influence the scheduling of these immense resources is severely limited. Therefore, the GPU is widely considered to be only suitable as an externally controlled co‐processor for homogeneous workloads which greatly restricts the potential applications of GPU computing. To address this issue, we present a new method to achieve fine‐grained priority scheduling on the GPU: hierarchical bucket queuing. By carefully distributing the workload among multiple queues and efficiently deciding which queue to draw work from next, we enable a variety of scheduling strategies. These strategies include fair‐scheduling, earliest‐deadline‐first scheduling and user‐defined dynamic priority scheduling. In a comparison with a sorting‐based approach, we reveal the advantages of hierarchical bucket queuing over previous work. Finally, we demonstrate the benefits of using priority scheduling in real‐world applications by example of path tracing and foveated micropolygon rendering.

We present a system which automatically transfers printed technical documentation, such as handbooks, to three-dimensional Augmented Reality. Our system identifies the most frequent forms of instructions found in printed documentation, such as image sequences, explosion diagrams, textual annotations and arrows indicating motion. The analysis of the printed documentation works automatically, with minimal user input. The system only requires the documentation itself and a CAD model or 3D scan of the object described in the documentation. The output is a fully interactive Augmented Reality application, presenting the information from the printed documentation in 3D, registered to the real object.

Abstract We present an approach for the automatic generation, interactive exploration and real‐time modification of disassembly procedures for complex, multipartite CAD data sets. In order to lift the performance barriers prohibiting interactive disassembly planning, we run a detailed analysis on the input model to identify recurring part constellations and efficiently determine blocked part motions in parallel on the GPU. Building on the extracted information, we present an interface for computing and editing extensive disassembly sequences in real‐time while considering user‐defined constraints and avoiding unstable configurations. To evaluate the performance of our C++/CUDA implementation, we use a variety of openly available CAD data sets, ranging from simple to highly complex. In contrast to previous approaches, our work enables interactive disassembly planning for objects which consist of several thousand parts and require cascaded translations during part removal.

In this paper, we present Whippletree, a novel approach to scheduling dynamic, irregular workloads on the GPU. We introduce a new programming model which offers the simplicity and expressiveness of task-based parallelism while retaining all aspects of the multi-level execution hierarchy essential to unlocking the full potential of a modern GPU. At the same time, our programming model lends itself to efficient implementation on the SIMD-based architecture typical of a current GPU. We demonstrate the practical utility of our model by providing a reference implementation on top of current CUDA hardware. Furthermore, we show that our model compares favorably to traditional approaches in terms of both performance as well as the range of applications that can be covered. We demonstrate the benefits of our model for recursive Reyes rendering, procedural geometry generation and volume rendering with concurrent irradiance caching.

In this paper we present Softshell, a novel execution model for devices composed of multiple processing cores operating in a single instruction, multiple data fashion, such as graphics processing units (GPUs). The Softshell model is intuitive and more flexible than the kernel-based adaption of the stream processing model, which is currently the dominant model for general purpose GPU computation. Using the Softshell model, algorithms with a relatively low local degree of parallelism can execute efficiently on massively parallel architectures. Softshell has the following distinct advantages: ( 1 ) work can be dynamically issued directly on the device, eliminating the need for synchronization with an external source, i.e ., the CPU; ( 2 ) its three-tier dynamic scheduler supports arbitrary scheduling strategies, including dynamic priorities and real-time scheduling; and ( 3 ) the user can influence, pause, and cancel work already submitted for parallel execution. The Softshell processing model thus brings capabilities to GPU architectures that were previously only known from operating-system designs and reserved for CPU programming. As a proof of our claims, we present a publicly available implementation of the Softshell processing model realized on top of CUDA. The benchmarks of this implementation demonstrate that our processing model is easy to use and also performs substantially better than the state-of-the-art kernel-based processing model for problems that have been difficult to parallelize in the past.