Taming 3DGS: High-Quality Radiance Fields with Limited Resources

3D Gaussian Splatting (3DGS) has transformed novel-view synthesis with its fast, interpretable, and high-fidelity rendering. However, its resource requirements limit its usability. Especially on constrained devices, training performance degrades quickly and often cannot complete due to excessive memory consumption of the model. The method converges with an indefinite number of Gaussians -- many of them redundant -- making rendering unnecessarily slow and preventing its usage in downstream tasks that expect fixed-size inputs. To address these issues, we tackle the challenges of training and rendering 3DGS models on a budget. We use a guided, purely constructive densification process that steers densification toward Gaussians that raise the reconstruction quality. Model size continuously increases in a controlled manner towards an exact budget, using score-based densification of Gaussians with training-time priors that measure their contribution. We further address training speed obstacles: following a careful analysis of 3DGS' original pipeline, we derive faster, numerically equivalent solutions for gradient computation and attribute updates, including an alternative parallelization for efficient backpropagation. We also propose quality-preserving approximations where suitable to reduce training time even further. Taken together, these enhancements yield a robust, scalable solution with reduced training times, lower compute and memory requirements, and high quality. Our evaluation shows that in a budgeted setting, we obtain competitive quality metrics with 3DGS while achieving a 4--5x reduction in both model size and training time. With more generous budgets, our measured quality surpasses theirs. These advances open the door for novel-view synthesis in constrained environments, e.g., mobile devices.

StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering

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 view-space 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.

LAENeRF: Local Appearance Editing for Neural Radiance Fields

Due to the omnipresence of Neural Radiance Fields (NeRFs), the interest towards editable implicit 3D representations has surged over the last years. However, editing implicit or hybrid representations as used for NeRFs is difficult due to the entanglement of appearance and geometry encoded in the model parameters. Despite these challenges, recent research has shown first promising steps towards photorealistic and non-photorealistic appearance edits. The main open issues of related work include limited interactivity, a lack of support for local edits and large memory requirements, rendering them less useful in practice. We address these limitations with LAENeRF, a unified framework for photorealistic and non-photorealistic appearance editing of NeRFs. To tackle local editing, we leverage a voxel grid as starting point for region selection. We learn a mapping from expected ray terminations to final output color, which can optionally be supervised by a style loss, resulting in a framework which can perform photorealistic and non-photorealistic appearance editing of selected regions. Relying on a single point per ray for our mapping, we limit memory requirements and enable fast optimization. To guarantee interactivity, we compose the output color using a set of learned, modifiable base colors, composed with additive layer mixing. Compared to concurrent work, LAENeRF enables recoloring and stylization while keeping processing time low. Furthermore, we demonstrate that our approach surpasses baseline methods both quantitatively and qualitatively.

Analyzing the Internals of Neural Radiance Fields

Modern Neural Radiance Fields (NeRFs) learn a mapping from position to volumetric density via proposal network samplers. In contrast to the coarse-to-fine sampling approach with two NeRFs, this offers significant potential for speedups using lower network capacity as the task of mapping spatial coordinates to volumetric density involves no view-dependent effects and is thus much easier to learn. Given that most of the network capacity is utilized to estimate radiance, NeRFs could store valuable density information in their parameters or their deep features. To this end, we take one step back and analyze large, trained ReLU-MLPs used in coarse-to-fine sampling. We find that trained NeRFs, Mip-NeRFs and proposal network samplers map samples with high density to local minima along a ray in activation feature space. We show how these large MLPs can be accelerated by transforming the intermediate activations to a weight estimate, without any modifications to the parameters post-optimization. With our approach, we can reduce the computational requirements of trained NeRFs by up to 50% with only a slight hit in rendering quality and no changes to the training protocol or architecture. We evaluate our approach on a variety of architectures and datasets, showing that our proposition holds in various settings.

Gradient-based Weight Density Balancing for Robust Dynamic Sparse Training

Training a sparse neural network from scratch requires optimizing connections at the same time as the weights themselves. Typically, the weights are redistributed after a predefined number of weight updates, removing a fraction of the parameters of each layer and inserting them at different locations in the same layers. The density of each layer is determined using heuristics, often purely based on the size of the parameter tensor. While the connections per layer are optimized multiple times during training, the density of each layer remains constant. This leaves great unrealized potential, especially in scenarios with a high sparsity of 90% and more. We propose Global Gradient-based Redistribution, a technique which distributes weights across all layers - adding more weights to the layers that need them most. Our evaluation shows that our approach is less prone to unbalanced weight distribution at initialization than previous work and that it is able to find better performing sparse subnetworks at very high sparsity levels.

MotionDeltaCNN: Sparse CNN Inference of Frame Differences in Moving Camera Videos

Convolutional neural network inference on video input is computationally expensive and has high memory bandwidth requirements. Recently, researchers managed to reduce the cost of processing upcoming frames by only processing pixels that changed significantly. Using sparse convolutions, the sparsity of frame differences can be translated to speedups on current inference devices. However, previous work was relying on static cameras. Moving cameras add new challenges in how to fuse newly unveiled image regions with already processed regions efficiently to minimize the update rate - without increasing memory overhead and without knowing the camera extrinsics of future frames. In this work, we propose MotionDeltaCNN, a CNN framework that supports moving cameras and variable resolution input. We propose a spherical buffer which enables seamless fusion of newly unveiled regions and previously processed regions - without increasing the memory footprint. Our evaluations show that we outperform previous work significantly by explicitly adding support for moving camera input.

AdaNeRF: Adaptive Sampling for Real-time Rendering of Neural Radiance Fields

Novel view synthesis has recently been revolutionized by learning neural radiance fields directly from sparse observations. However, rendering images with this new paradigm is slow due to the fact that an accurate quadrature of the volume rendering equation requires a large number of samples for each ray. Previous work has mainly focused on speeding up the network evaluations that are associated with each sample point, e.g., via caching of radiance values into explicit spatial data structures, but this comes at the expense of model compactness. In this paper, we propose a novel dual-network architecture that takes an orthogonal direction by learning how to best reduce the number of required sample points. To this end, we split our network into a sampling and shading network that are jointly trained. Our training scheme employs fixed sample positions along each ray, and incrementally introduces sparsity throughout training to achieve high quality even at low sample counts. After fine-tuning with the target number of samples, the resulting compact neural representation can be rendered in real-time. Our experiments demonstrate that our approach outperforms concurrent compact neural representations in terms of quality and frame rate and performs on par with highly efficient hybrid representations. Code and supplementary material is available at https://thomasneff.github.io/adanerf.

DeltaCNN: End-to-End CNN Inference of Sparse Frame Differences in Videos

Convolutional neural network inference on video data requires powerful hardware for real-time processing. Given the inherent coherence across consecutive frames, large parts of a video typically change little. By skipping identical image regions and truncating insignificant pixel updates, computational redundancy can in theory be reduced significantly. However, these theoretical savings have been difficult to translate into practice, as sparse updates hamper computational consistency and memory access coherence; which are key for efficiency on real hardware. With DeltaCNN, we present a sparse convolutional neural network framework that enables sparse frame-by-frame updates to accelerate video inference in practice. We provide sparse implementations for all typical CNN layers and propagate sparse feature updates end-to-end - without accumulating errors over time. DeltaCNN is applicable to all convolutional neural networks without retraining. To the best of our knowledge, we are the first to significantly outperform the dense reference, cuDNN, in practical settings, achieving speedups of up to 7x with only marginal differences in accuracy.

DONeRF: Towards Real-Time Rendering of Neural Radiance Fields using Depth Oracle Networks

The recent research explosion around implicit neural representations, such as NeRF, shows that there is immense potential for implicitly storing high-quality scene and lighting information in neural networks. However, one major limitation preventing the use of NeRF in interactive and real-time rendering applications is the prohibitive computational cost of excessive network evaluations along each view ray, requiring dozens of petaFLOPS when aiming for real-time rendering on consumer hardware. In this work, we take a step towards bringing neural representations closer to practical rendering of synthetic content in interactive and real-time applications, such as games and virtual reality. We show that the number of samples required for each view ray can be significantly reduced when local samples are placed around surfaces in the scene. To this end, we propose a depth oracle network, which predicts ray sample locations for each view ray with a single network evaluation. We show that using a classification network around logarithmically discretized and spherically warped depth values is essential to encode surface locations rather than directly estimating depth. The combination of these techniques leads to DONeRF, a dual network design with a depth oracle network as a first step and a locally sampled shading network for ray accumulation. With our design, we reduce the inference costs by up to 48x compared to NeRF. Using an off-the-shelf inference API in combination with simple compute kernels, we are the first to render raymarching-based neural representations at interactive frame rates (15 frames per second at 800x800) on a single GPU. At the same time, since we focus on the important parts of the scene around surfaces, we achieve equal or better quality compared to NeRF to enable interactive high-quality rendering.

UNOC: Understanding Occlusion for Embodied Presence in Virtual Reality

Tracking body and hand motions in the 3D space is essential for social and self-presence in augmented and virtual environments. Unlike the popular 3D pose estimation setting, the problem is often formulated as inside-out tracking based on embodied perception (e.g., egocentric cameras, handheld sensors). In this paper, we propose a new data-driven framework for inside-out body tracking, targeting challenges of omnipresent occlusions in optimization-based methods (e.g., inverse kinematics solvers). We first collect a large-scale motion capture dataset with both body and finger motions using optical markers and inertial sensors. This dataset focuses on social scenarios and captures ground truth poses under self-occlusions and body-hand interactions. We then simulate the occlusion patterns in head-mounted camera views on the captured ground truth using a ray casting algorithm and learn a deep neural network to infer the occluded body parts. In the experiments, we show that our method is able to generate high-fidelity embodied poses by applying the proposed method on the task of real-time inside-out body tracking, finger motion synthesis, and 3-point inverse kinematics.