QS-FedNeRF: Quantized transmission and device selections federated neural radiance fields for edge intelligence

This paper studies the task of SatStreet-view synthesis, which aims to render photorealistic street-view panorama images and videos given a satellite image and specified camera positions or trajectories. Our approach involves learning a satellite image conditioned neural radiance field from paired images captured from both satellite and street viewpoints, which comes to be a challenging learning problem due to the sparse-view nature and the extremely large viewpoint changes between satellite and street-view images. We tackle the challenges based on a task-specific observation that street-view specific elements, including the sky and illumination effects, are only visible in street-view panoramas, and present a novel approach, Sat2Density++, to accomplish the goal of photo-realistic street-view panorama rendering by modeling these street-view specific elements in neural networks. In the experiments, our method is evaluated on both urban and suburban scene datasets, demonstrating that Sat2Density++ is capable of rendering photorealistic street-view panoramas that are consistent across multiple views and faithful to the satellite image.

3D scene reconstruction serves as a fundamental technology with widespread applications in virtual reality, structural inspection, and robotic systems. While recent advances in 3D Gaussian Splatting have significantly enhanced scene reconstruction capabilities, the performance of such methods remains suboptimal when applied to highly reflective environments. To overcome this limitation, we introduce ARS-GS, a novel framework that integrates Anisotropic Spherical Gaussian reflection modeling and spherical harmonics diffuse approximation into a physically based rendering pipeline. This architecture incorporates a skip connection between the Anisotropic Spherical Gaussian module and the Gaussian primitives, effectively preserving surface details while maintaining computational efficiency. Comprehensive experimental evaluations validate the efficacy of ARS-GS across multiple datasets. Specifically, our method establishes new state-of-the-art quantitative benchmarks, achieving a peak signal-to-noise ratio of 38.30 and a structural similarity index measure of 0.997 on the neural radiance fields synthetic dataset, alongside a peak signal-to-noise ratio of 46.31 on the Gloss Blender dataset. Furthermore, on the challenging reflective neural radiance fields real-world dataset, our approach secures the highest peak signal-to-noise ratio scores, highlighted by a metric of 26.26 on the Sedan scene. The proposed method also substantially reduces perceptual errors, yielding a learned perceptual image patch similarity as low as 0.204, thereby consistently outperforming existing techniques in the reconstruction of highly specular surfaces with superior geometric fidelity.

Unsupervised Large-Cone-Angle Artifact Suppression for CBCT Reconstruction via Neural Radiance Fields

ABSTRACT 3D reconstruction and novel view synthesis tasks play a crucial role in fields such as digital twins and spatial perception, providing essential data support for these applications. However, in low-texture and textureless scenes, these tasks face significant challenges due to the absence of prominent visual features. In recent years, 3D Gaussian Splatting (3DGS) technology has demonstrated superior performance compared to traditional 3D reconstruction methods and Neural Radiance Fields (NeRF) based novel view synthesis techniques, owing to its explicit representation capabilities and efficient real-time realistic rendering performance. Nevertheless, in low-textured scenes, 3DGS still inevitably suffers from insufficient geometric constraints due to a significant reduction in feature density, which consequently impacts reconstruction accuracy and rendering quality. This paper explores the use of globally consistent monocular depth initialization in 3DGS to solve the 3D reconstruction task of scenes involving large areas of weak texture. A graph neural network (GNN) depth prediction module supervised by SfM sparse point cloud was designed, and the attention coefficient was adjusted according to node importance by combining graph attention mechanism. The depth of SfM sparse point cloud was propagated to all pixels, eliminating global scale drift, enhancing scene structure consistency, and providing globally consistent high-quality depth data for 3DGS initialization. In addition to photometric, depth, and normal constraints, Fast Point Feature Histogram (FPFH) features suitable for describing the geometric features of complex scenes have been added for geometric regularization to further ensure geometric consistency. Numerous experiments on our datasets have shown that our method outperforms the original 3DGS in low-textured scenes, achieving an improvement of more than 0.295 dB in Peak Signal-to-Noise Ratio (PSNR).

Neural rendering has been the disruptive technology in the creation of very realistic content of the visual multimedia production that the computer graphics and deep learning have substituted. This paper examines the neural rendering systems that can be trained to produce hyper-realistic artistic images through the acquisition of the complex representations of scenes based on multi-view image representations. The suggested architecture compiles the neural radiance field modeling, deep neural networks as well as volumetric rendering to reproduce detailed three-dimensional scenes as well as produce photorealistic images in new perspectives. Multi-view data acquisition, neural feature encoding, and radiance field estimation are the system architecture elements based on deep learning models that capture geometry, lighting, texture and color interaction within a scene. Experimental analysis of neural rendering methods has shown that they render visual fidelity, geometric consistency and rendering realism by a wide margin than the standard computer graphics pipelines. The quantitative investigation of the measures of the quality of rendering, such as the similarity index of the structure, the perceptual realism scores, and the reconstruction accuracy, reveals the significant progress of visual detail and scene modeling.

Modeling the wireless radiance field (WRF) is fundamental to modern communication systems, enabling key tasks such as localization, sensing, and channel estimation. Traditional approaches, which rely on empirical formulas or physical simulations, often suffer from limited accuracy or require strong scene priors. Recent neural radiance field (NeRF)-based methods improve reconstruction fidelity through differentiable volumetric rendering, but their reliance on computationally expensive multilayer perceptron (MLP) queries hinders real-time deployment. To overcome these challenges, we introduce Gaussian splatting (GS) to the wireless domain, leveraging its efficiency in modeling optical radiance fields to enable compact and accurate WRF reconstruction. Specifically, we propose SwiftWRF, a deformable 2D Gaussian splatting framework that synthesizes WRF spectra at arbitrary positions under single-sided transceiver mobility. SwiftWRF employs CUDA-accelerated rasterization to render spectra at over 100k FPS and uses the lightweight MLP to model the deformation of 2D Gaussians, effectively capturing mobility-induced WRF variations. In addition to novel spectrum synthesis, the efficacy of SwiftWRF is further underscored in its applications in angle-of-arrival (AoA) and received signal strength indicator (RSSI) prediction. Experiments conducted on both real-world and synthetic indoor scenes demonstrate that SwiftWRF can reconstruct WRF spectra up to 500x faster than existing state-of-the-art methods, while significantly enhancing its signal quality.

This study presents a comparative evaluation of Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) for 3D reconstruction in indoor and outdoor environments. High-quality 3D models are significant in a range of applications, from forensic investigations to cultural heritage preservation, architecture, and robotics, where detail accuracy and minimal noise are crucial. Leveraging continuous video footage captured with a stabilized full-frame camera setup, this research examines both algorithms across indoor and outdoor environments using consistent datasets. Key assessment criteria include reconstruction noise, detail preservation, and processing time. The results reveal that while both approaches generate high fidelity reconstructions, 3DGS outperforms NeRF in computational efficiency and noise reduction. These insights provide valuable guidance for selecting suitable reconstruction techniques across different professional domains. Due to the controlled scope and limited number of test scenes, the findings should be interpreted as indicative rather than statistically generalizable, serving primarily as a practical, application-oriented comparison.

Recent advances in leveraging large-scale Internet photo collections for 3D reconstruction have enabled immersive virtual exploration of landmarks and historic sites worldwide. However, existing methods primarily focus on visual appearance reconstruction, often overlooking the interactive semantic understanding of these 3D scenes (e.g., identifying specific building parts or scene details), which remains largely confined to browsing static text-image pairs. Therefore, can we draw inspiration from 3D in-the-wild reconstruction techniques and use unconstrained photo collections to create an immersive approach for comprehensive 3D scene understanding beyond mere visual appearance? To this end, we extend language embedded 3D Gaussian splatting (3DGS) and propose a novel framework for open-vocabulary scene understanding from unconstrained photo collections. Specifically, we first render multiple appearance images from the same viewpoint as the unconstrained image with the reconstructed radiance field, then extract multi-appearance CLIP features and two types of language feature uncertainty maps-transient and appearance uncertainty-derived from the multi-appearance features to guide the subsequent optimization process. Next, we propose a transient uncertainty-aware autoencoder, a multi-appearance language field 3DGS representation, and a post-ensemble strategy to effectively compress, learn, and fuse language features from multiple appearances. Finally, to quantitatively evaluate our method, we introduce PT-OVS, a new benchmark dataset for assessing open-vocabulary segmentation performance on unconstrained photo collections. Experimental results show that our method outperforms existing methods, delivering accurate open-vocabulary segmentation and enabling applications such as interactive roaming with open-vocabulary queries, architectural style pattern recognition, and 3D scene editing.

As the mussel farming industry grows, the demand for advanced monitoring and management solutions intensifies. Traditional methods rely heavily on on-site observations and frequent boat trips to monitor buoy flotation and other operational elements, often resulting in limited and sporadic assessments that can hinder the decision-making process. This article proposes using 3D reconstruction techniques to reconstruct mussel farm scenes from the vessel-captured video footage, allowing for holistic visualizations from various perspectives and enabling comprehensive post analysis of the mussel farm dynamics. While previous efforts relied on traditional Structure from Motion and multiview techniques for mussel farm scene reconstruction, they often struggled to capture fine details and faced challenges with reflective water surfaces due to their reliance on local features. In this work, we are the first to explore and extend the capabilities of neural radiance field (NeRF) for mussel farm reconstruction. To overcome the practical challenges of this unique environment, we propose a multi-NeRF framework with region-specific modeling, enabling the capture of both the global scene and finer details of key elements such as buoys. Furthermore, we introduce a geometry regularization method to improve the planar reconstruction of the water surface. Our results demonstrate significant advancements in 3D reconstruction quality over previous methods, particularly in mesh completeness and the precise handling of specular and diffuse texture details while also enabling realistic novel view synthesis. These advancements, designed particularly for mussel farm applications, can contribute to its intelligent monitoring and management by providing a comprehensive understanding of the farm's geometry and dynamics, ultimately facilitating more informed decision-making processes.

SS-NeRF: Shine-sphere rendering for neural radiance fields

PGgraf: Pose-Guided generative radiance field for novel-views on X-ray

HiSURF: Hierarchical semantic-guided unified radiance field for generalizing across unseen scenes

Abstract In recent years, the development of Neural Radiance Fields (NeRF) technology has brought significant breakthroughs to the field of deblurring in 3D reconstruction. Although existing methods can recover relatively clear 3D scenes from blurred images, their effectiveness remains limited when handling severely blurred inputs. To address this challenge, we propose EDD-NeRF, an innovative framework that explicitly reconstructs complex camera motion trajectories during exposure by employing cubic B-spline interpolation on the SE(3) manifold. By integrating event camera data, depth priors, and a novel MLP architecture, our approach achieves high-quality reconstruction of severely blurred scenes. Evaluations demonstrate that our method significantly outperforms the original NeRF and other deblurring NeRF techniques in reconstruction quality.

Neural Radiance Fields (NeRF) have shown remarkable capabilities for photorealistic novel view synthesis. One major deficiency of NeRF is that dense inputs are typically required, and the rendering quality will drop drastically given sparse inputs. In this paper, we highlight the effectiveness of rendered semantics from dense novel views, and show that rendered semantics can be treated as a more robust form of augmented data than rendered RGB. Our method enhances NeRF’s performance by incorporating guidance derived from the rendered semantics. The rendered semantic guidance encompasses two levels: the supervision level and the feature level. The supervision-level guidance incorporates a bi-directional verification module that decides the validity of each rendered semantic label, while the feature-level guidance integrates a learnable codebook that encodes semantic-aware information, which is queried by each point via the attention mechanism to obtain semanticrelevant predictions. The overall semantic guidance is embedded into a self-improved pipeline.We also introduce a more challenging sparse-input indoor benchmark, where the number of inputs is limited to as few as 6. Experiments demonstrate the effectiveness of our method and it exhibits superior performance compared to existing approaches.

The acquisition of high-quality three-dimensional (3D) models of cultural relics often relies on expensive scanning equipment or multi-view image capture, which limits large-scale deployment in real-world heritage conservation scenarios. Large-scale water impoundment in the Three Gorges region has resulted in the permanent submergence of numerous cultural relics and archaeological remains. For many of these artifacts, only a single two-dimensional image remains as the sole visual record, posing significant challenges for reconstructing their original three-dimensional geometry and appearance. This limitation renders traditional multi-view reconstruction and physical scanning methods infeasible. To address this challenge, we propose a generative framework for reconstructing high-fidelity 3D digital models of Chinese Three Gorges cultural relics from a single two-dimensional (2D) image. Building upon recent advances in generative 3D representation learning, the proposed method adopts a transformer-based image-to-triplane architecture to infer an implicit 3D representation directly from a single RGB image. A vision transformer encoder is employed to extract global and local visual features, which are subsequently projected into a compact triplane representation through a cross-attention-based decoder. The reconstructed triplane features are further decoded by a neural radiance field (NeRF) to synthesize dense geometry and appearance, enabling accurate mesh extraction and novel-view rendering. To enhance robustness under in-the-wild conditions, the model implicitly estimates camera parameters during inference without relying on explicit calibration information. The proposed method is evaluated on a dataset of Chinese Three Gorges cultural relics, covering diverse artifact categories and visual styles. Experimental results demonstrate that the proposed framework is capable of producing structurally coherent and visually consistent 3D reconstructions from a single image, effectively preserving key morphological characteristics of cultural relics under limited data conditions. Compared with existing single-image and multi-view reconstruction baselines, the proposed framework exhibits better reconstruction accuracy, visual consistency, and generalization capability. This study provides an efficient and scalable solution for the digital reconstruction of cultural relics and offers a practical pathway for large-scale 3D digitization of heritage artifacts from archival images. This work provides a practical solution for the digital reconstruction of submerged heritage artifacts and contributes to the application of generative 3D modeling techniques in cultural heritage preservation and restoration.

Aerial remote sensing efficiently surveys large areas, but accurate direct object-level measurement remains difficult in complex natural scenes. Advancements in 3D computer vision, particularly radiance field representations such as NeRF and 3D Gaussian splatting, can improve reconstruction fidelity from posed imagery. Nevertheless, direct aerial measurement of important attributes like tree diameter at breast height (DBH) remains challenging. Trunks in aerial forest scans are distant and sparsely observed in image views; at typical operating altitudes, stems may span only a few pixels. With these constraints, conventional reconstruction methods have inaccurate breast-height trunk geometry. TreeDGS is an aerial image reconstruction method that uses 3D Gaussian splatting as a continuous scene representation for trunk measurement. After SfM–MVS initialization and Gaussian optimization, we extract a dense point set from the Gaussian field using RaDe-GS’s depth-aware cumulative-opacity integration and associate each sample with a multi-view opacity reliability score. Then, we isolate trunk points and estimate DBH using opacity-weighted solid-circle fitting. Evaluated on 10 plots with field-measured DBH, TreeDGS reaches 4.79 cm RMSE (about 2.6 pixels at this GSD) and outperforms a LiDAR baseline (7.66 cm RMSE). This shows that TreeDGS can enable accurate, low-cost aerial DBH measurement.

The proliferation of immersive 3D web applications, from e-commerce product viewers to virtual real estate tours, has created a critical need for high-quality, real-time rendering directly within the browser. Neural radiance fields (NeRF) offer unprecedented photorealism but are hamstrung by immense computational demands, making their deployment on resource-constrained web platforms a significant web engineering challenge. The core bottleneck is NeRF’s reliance on dense point sampling for volume rendering. This paper introduces a novel framework that directly tackles this challenge through a pioneering adaptive sampling technique powered by reinforcement learning. We name this framework PPO-NeRF. It integrates the rapid training capabilities of Instant-NGP’s hash encoding with an agent trained via proximal policy optimization (PPO). This agent learns to adaptively predict the minimal set of crucial sample points along each camera ray, dynamically pruning computationally redundant samples to optimize rendering specifically for web-based, real-time scenarios. Experimental results demonstrate that PPO-NeRF significantly lowers the barrier to web deployment. Compared to the original NeRF, it reduces training time by approximately 73.63%, enabling faster content iteration for web developers. More critically, our adaptive sampling slashes rendering time by approximately 44.7% and VRAM usage by approximately 29.9%, while maintaining comparable visual fidelity. These gains directly translate to faster load times, smoother user interaction, and broader device compatibility. In conclusion, PPO-NeRF provides a practical solution to NeRF’s long-standing performance bottlenecks, establishing a viable pathway for deploying high-fidelity, interactive 3D experiences at scale across the modern web.

The proliferation of Unmanned Aerial Vehicles (UAVs) has significantly advanced large-scale 3D data acquisition, paving the way for Digital Twin applications. While Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) now enable exceptional photorealism, editing these vast neural representations remains a significant challenge. This survey systematically analyzes large-scale scene editing from UAV imagery, specifically focusing on the architectural dichotomy between implicit (NeRF) and explicit (3DGS) representations. We introduce novel taxonomies to deconstruct their core mechanisms and critically evaluate why explicit representations are currently emerging as the more viable path for interactive, city-scale manipulation. Furthermore, we identify key bottlenecks—such as handling oblique views and variable atmospheric conditions—and outline future research directions.A comprehensive and regularly updated collection of the surveyed papers and resources is available at our project website: https://github.com/ruigong335-creator/A-Survey-on-Aerial-Scene-Editing .

Neural Radiance Fields (NeRF) have significantly advanced photorealistic novel view synthesis. Recently, 3D Gaus sian Splatting has emerged as a promising technique with faster training and rendering speeds. However, both methods rely heavily on clear images and precise camera poses, limiting performance under motion blur. To address this, we introduce Event-Informed 3D Deblur Reconstruction with Gaussian Splat ting(EiGS), a novel approach leveraging event camera data to enhance 3D Gaussian Splatting, improving sharpness and clarity in scenes affected by motion blur. Our method employs an Adaptive Deviation Estimator to learn Gaussian center shifts as the inverse of complex camera jitter, enabling simulation of motion blur during training. A motion consistency loss ensures global coherence in Gaussian displacements, while Blurriness and Event Integration Losses guide the model toward precise 3D representations. Extensive experiments demonstrate superior sharpness and real-time rendering capabilities compared to existing methods, with ablation studies validating the effectiveness of our components in robust, high-quality reconstruction for complex static scenes.