View Synthesis Method Research Articles

FoV-NeRF: Foveated Neural Radiance Fields for Virtual Reality.

Virtual Reality (VR) is becoming ubiquitous with the rise of consumer displays and commercial VR platforms. Such displays require low latency and high quality rendering of synthetic imagery with reduced compute overheads. Recent advances in neural rendering showed promise of unlocking new possibilities in 3D computer graphics via image-based representations of virtual or physical environments. Specifically, the neural radiance fields (NeRF) demonstrated that photo-realistic quality and continuous view changes of 3D scenes can be achieved without loss of view-dependent effects. While NeRF can significantly benefit rendering for VR applications, it faces unique challenges posed by high field-of-view, high resolution, and stereoscopic/egocentric viewing, typically causing low quality and high latency of the rendered images. In VR, this not only harms the interaction experience but may also cause sickness. To tackle these problems toward six-degrees-of-freedom, egocentric, and stereo NeRF in VR, we present the first gaze-contingent 3D neural representation and view synthesis method. We incorporate the human psychophysics of visual- and stereo-acuity into an egocentric neural representation of 3D scenery. We then jointly optimize the latency/performance and visual quality while mutually bridging human perception and neural scene synthesis to achieve perceptually high-quality immersive interaction. We conducted both objective analysis and subjective studies to evaluate the effectiveness of our approach. We find that our method significantly reduces latency (up to 99% time reduction compared with NeRF) without loss of high-fidelity rendering (perceptually identical to full-resolution ground truth). The presented approach may serve as the first step toward future VR/AR systems that capture, teleport, and visualize remote environments in real-time.

SNeRF

This paper presents a stylized novel view synthesis method. Applying state-of-the-art stylization methods to novel views frame by frame often causes jittering artifacts due to the lack of cross-view consistency. Therefore, this paper investigates 3D scene stylization that provides a strong inductive bias for consistent novel view synthesis. Specifically, we adopt the emerging neural radiance fields (NeRF) as our choice of 3D scene representation for their capability to render high-quality novel views for a variety of scenes. However, as rendering a novel view from a NeRF requires a large number of samples, training a stylized NeRF requires a large amount of GPU memory that goes beyond an off-the-shelf GPU capacity. We introduce a new training method to address this problem by alternating the NeRF and stylization optimization steps. Such a method enables us to make full use of our hardware memory capacity to both generate images at higher resolution and adopt more expressive image style transfer methods. Our experiments show that our method produces stylized NeRFs for a wide range of content, including indoor, outdoor and dynamic scenes, and synthesizes high-quality novel views with cross-view consistency.

Do We Need a New Large-Scale Quality Assessment Database for Generative Inpainting Based 3D View Synthesis? (Student Abstract)

The advancement in Image-to-Image translation techniques using generative Deep Learning-based approaches has shown promising results for the challenging task of inpainting-based 3D view synthesis. At the same time, even the current 3D view synthesis methods often create distorted structures or blurry textures inconsistent with surrounding areas. We analyzed the recently proposed algorithms for inpainting-based 3D view synthesis and observed that these algorithms no longer produce stretching and black holes. However, the existing databases such as IETR, IRCCyN, and IVY have 3D-generated views with these artifacts. This observation suggests that the existing 3D view synthesis quality assessment algorithms can not judge the quality of most recent 3D synthesized views. With this view, through this abstract, we analyze the need for a new large-scale database and a new perceptual quality metric oriented for 3D views using a test dataset.

Dense view synthesis for three-dimensional light-field display based on scene geometric reconstruction

The three-dimensional light-field display is one of the most promising display technologies and has achieved an impressive improvement in past years. To generate real-world dense views efficiently, many virtual view synthesis approaches have been proposed, which take sparse 2D images as input and synthesize dense virtual views. However, in the occlusion areas, the generated results of previous methods are not satisfactory, which usually contain errors and blurs because the occlusion relation cannot​ be recovered correctly. Here, a dense view synthesis method for three-dimensional light-field display based on scene geometric reconstruction is presented, which can synthesize high-quality arbitrary virtual views with correct occlusions. The scene geometric model is firstly reconstructed with several depth maps captured by RGB-D cameras. The geometric model can provide the correct occlusion relation and select the non-occlusion regions as the input data. To relieve the effect of the depth map noise and improve the quality further, the input views semantic feature maps are extracted and reprojected to the virtual view, which encodes the captured multi-view information and represents the virtual view implicitly. The virtual view image can be generated by rendering the reprojected encoded feature map. Compared with other view synthesis approaches, the quality of the synthesized view is significantly improved with our proposed method, especially in the occlusion areas. The experiments are carried out on a 27 inches 3D light-field display device, and the results demonstrate the effectiveness of our proposed method.

Virtual view synthesis for 3D light-field display based on feature reprojection and fusion

Three-dimensional light field display has achieved an impressive result and is a great potential display method. To provide real-world dense views efficiently, many virtual view synthesis approaches have been proposed. However, previous methods are not robust enough when the input view poses with a wider baseline or large rotation. Here, a virtual view synthesis method based on feature reprojection and fusion is presented, which can robustly generate high-quality dense virtual views for 3D light-field display. The target view depth map is generated with our proposed unsupervised network, which provides the 3D scene distribution priori. Then input views’ feature maps are reprojected to the target view position. To fuse these reprojected feature maps effectively, a feature fusion Transformer network is designed, which encodes all input views’ information and generates a fused feature map to represent the target view implicitly. The target view color map can be synthesized by rendering the fused feature map. Experimental results demonstrate the effectiveness of the proposed method, which shows outstanding performances when the input view poses are arranged unregularly. We believe the proposed method relieves the requirements of input sparse views acquisition and contributes to the wide applications of 3D light-field display in the future.

Machine Learning-Based View Synthesis in Fourier Lightfield Microscopy.

Current interest in Fourier lightfield microscopy is increasing, due to its ability to acquire 3D images of thick dynamic samples. This technique is based on simultaneously capturing, in a single shot, and with a monocular setup, a number of orthographic perspective views of 3D microscopic samples. An essential feature of Fourier lightfield microscopy is that the number of acquired views is low, due to the trade-off relationship existing between the number of views and their corresponding lateral resolution. Therefore, it is important to have a tool for the generation of a high number of synthesized view images, without compromising their lateral resolution. In this context we investigate here the use of a neural radiance field view synthesis method, originally developed for its use with macroscopic scenes acquired with a moving (or an array of static) digital camera(s), for its application to the images acquired with a Fourier lightfield microscope. The results obtained and presented in this paper are analyzed in terms of lateral resolution and of continuous and realistic parallax. We show that, in terms of these requirements, the proposed technique works efficiently in the case of the epi-illumination microscopy mode.

A neural refinement network for single image view synthesis

Recent years have seen an increasing interest in single image view synthesis. It remains however a challenging task due to the lack of comprehensive colour and depth information from different views. In this paper, we propose a novel view synthesis approach that incorporates a Neural Image Refinement Network (NIRN) and generates both depth and colour images for the target view in an end-to-end manner. The appearance of the colour image greatly benefits from the generated depth image as it provides an intermediate projection relationship for the object in the 3D world. Since the direct application of geometric projection mapping will result in empty regions and/or distortions, our approach proposes to embed a novel refinement network into the view synthesis pipeline for improved performance. Experimental results on three publicly available datasets demonstrate that our NIRN outperforms other state-of-the-art view synthesis methods.

Inference-Reconstruction Variational Autoencoder for Light Field Image Reconstruction.

Light field cameras can capture the radiance and direction of light rays by a single exposure, providing a new perspective to photography and 3D geometry perception. However, existing sub-aperture based light field cameras are limited by their sensor resolution to obtain high spatial and angular resolution images simultaneously. In this paper, we propose an inference-reconstruction variational autoencoder (IR-VAE) to reconstruct a dense light field image out of four corner reference views in a light field image. The proposed IR-VAE is comprised of one inference network and one reconstruction network, where the inference network infers novel views from existing reference views and viewpoint conditions, and the reconstruction network reconstructs novel views from a latent variable that contains the information of reference views, novel views, and viewpoints. The conditional latent variable in the inference network is regularized by the latent variable in the reconstruction network to facilitate information flow between the conditional latent variable and novel views. We also propose a statistic distance measurement dubbed the mean local maximum mean discrepancy (MLMMD) to enable the measurement of the statistic distance between two distributions with high-resolution latent variables, which can capture richer information than their low-resolution counterparts. Finally, we propose a viewpoint-dependent indirect view synthesis method to synthesize novel views more efficiently by leveraging adaptive convolution. Experimental results show that our proposed methods outperform state-of-the-art methods on different light field datasets.

Remote Sensing Novel View Synthesis With Implicit Multiplane Representations

Novel view synthesis of remote sensing scenes is of great significance for scene visualization, human-computer interaction, and various downstream applications. Despite the recent advances in computer graphics and photogrammetry technology, generating novel views is still challenging particularly for remote sensing images due to its high complexity, view sparsity and limited view-perspective variations. In this paper, we propose a novel remote sensing view synthesis method by leveraging the recent advances in implicit neural representations. Considering the overhead and far depth imaging of remote sensing images, we represent the 3D space by combining implicit multiplane images (MPI) representation and deep neural networks. The 3D scene is reconstructed under a self-supervised optimization paradigm through a differentiable multiplane renderer with multi-view input constraints. Images from any novel views thus can be freely rendered on the basis of the reconstructed model. As a by-product, the depth maps corresponding to the given viewpoint can be generated along with the rendering output. We refer to our method as Implicit Multiplane Images (ImMPI). To further improve the view synthesis under sparse-view inputs, we explore the learning-based initialization of remote sensing 3D scenes and proposed a neural network based Prior extractor to accelerate the optimization process. In addition, we propose a new dataset for remote sensing novel view synthesis with multi-view real-world google earth images. Extensive experiments demonstrate the superiority of the ImMPI over previous state-of-the-art methods in terms of reconstruction accuracy, visual fidelity, and time efficiency. Ablation experiments also suggest the effectiveness of our methodology design. Our dataset and code can be found at https://github.com/wyc-Chang/ImMPI

Attention Mechanism-Based Light-Field View Synthesis

The angular information of light lost in conventional images but preserved and stored in light-fields plays an instrumental role in many applications such as depth estimation, 3D reconstruction and post-capture refocusing. However, the limited angular resolution of light-field images due to the consumer hardware limitations is a major drawback in its widespread adoption. In this article, we present a novel deep learning-based light-field view synthesis method from a sparse set of input views. Our proposed method, an end-to-end trainable network, utilizes convolutional block attention modules to enhance the built-in depth image-based rendering. The proposed convolutional block attention module consists of two attention modules sequentially applied in the channel and spatial dimensions to focus the network on critical features. The proposed network architecture is a combination of three sub-networks, one for stereo feature extraction, another for disparity estimation, and the last for attention-based refinement. We present two schemes for the refinement network that perform equally well but differ in the number of parameters by 1.2% and execution time by 44%. Quantitative and qualitative results on four challenging real-world datasets show the superiority of the proposed method. Our proposed method shows considerable PSNR gains, around 1 dB compared to the state of the art and around 0.5 dB over our previous method LFVS-AM. In addition, ablation studies demonstrate the effectiveness of each module of the proposed method. Finally, the parallax edge precision-recall curve shows that our method better preserves parallax details.

TM-NET

We introduce TM-NET, a novel deep generative model for synthesizing textured meshes in a part-aware manner. Once trained, the network can generate novel textured meshes from scratch or predict textures for a given 3D mesh, without image guidance. Plausible and diverse textures can be generated for the same mesh part, while texture compatibility between parts in the same shape is achieved via conditional generation. Specifically, our method produces texture maps for individual shape parts, each as a deformable box, leading to a natural UV map with limited distortion. The network separately embeds part geometry (via a PartVAE) and part texture (via a TextureVAE) into their respective latent spaces, so as to facilitate learning texture probability distributions conditioned on geometry. We introduce a conditional autoregressive model for texture generation, which can be conditioned on both part geometry and textures already generated for other parts to achieve texture compatibility. To produce high-frequency texture details, our TextureVAE operates in a high-dimensional latent space via dictionary-based vector quantization. We also exploit transparencies in the texture as an effective means to model complex shape structures including topological details. Extensive experiments demonstrate the plausibility, quality, and diversity of the textures and geometries generated by our network, while avoiding inconsistency issues that are common to novel view synthesis methods.

Unsupervised virtual view synthesis from monocular video with generative adversarial warping

Virtual view synthesis from monocular video is challenging, as it aims to infer photorealistic views from single reference view. Previous work have achieved acceptable visual quality, however, are heavily relied on supervision information, such as depth or pristine virtual view, which are inadequate in practical application. In this paper, an unsupervised virtual view synthesis method is proposed to get rid of the supervision information. Firstly, it embed a spatiotemporal generative adversarial network into traditional depth-image-based rendering framework with no explicit depth information provided. Secondly, it utilized novel perceptual constraints without relying on pristine images, including the blind synthesized image quality metric and no-reference structure similarity. The entire framework is fully convolutional, producing hallucinated results in an end-to-end way. Particularly, the whole framework is independent of supervision information. Experimental results demonstrate that the proposed method produces pleasant virtual views in comparison with supervised methods, thereby can be beneficial to practical applications.

Combining Retargeting Quality and Depth Perception Measures for Quality Evaluation of Retargeted Stereopairs

Stereoscopic Image Retargeting (SIR) aims to adapt stereoscopic images and videos to 3D display devices with various aspect ratios by emphasizing the important content while retaining surrounding context with minimal visual distortion. To address the issue of SIR evaluation, this paper presents a new objective quality assessment method for retargeted stereopairs by combining image quality and depth perception measures. Specifically, the image quality measure is conducted between the source and retargeted intermediate views generated by the view synthesis method to characterize the geometric distortion and content loss of the retargeted stereopair, while several depth-aware features are extracted to measure the visual comfort/discomfort and depth sensation when human views a 3D scene. Then, the extracted features are integrated into an overall perceptual quality prediction. Experiment results on NBU SIRQA and SIRD databases verify the superiority of our method.

Fixing Defect of Photometric Loss for Self-Supervised Monocular Depth Estimation

View-synthesis-based methods have shown very promising results for the task of unsupervised depth estimation in single images. Most existing approaches synthesize a new image and employ it as the supervision signal for depth and pose prediction. There are two problems in these approaches: 1) There are many combinations of pose and depth that can synthesize a certain new image; therefore, reconstructing the depth and pose based on the view-synthesis method from only two images is an inherently ill-posed problem; 2) The model is trained under the photometric consistency assumption that the brightness or gradient is constant when applied to the video sequences. However, this assumption is easily violated in realistic scenes due to light changes, reflective surfaces and occlusions. To overcome the first drawback, we exploit the point cloud consistency constraint to eliminate ambiguity. To overcome the second drawback, we use threshold masks to filter dynamic and occluded points and introduce matching point constraints that implicitly encode the geometry relationship between two matched points to improve the precision of depth prediction. In addition, we employ epipolar constraints to compensate for the instability of the photometric error in textureless regions and varying illumination conditions. The experimental results on the KITTI, Cityscapes and NYUv2 datasets show that the method can improve the accuracy of depth prediction and enhance the robustness of the model in handling textureless regions and illumination changes. The code and data are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/XTUPRLAB/FixUnDepth</uri> .

Virtual view synthesis for 3D light-field display based on scene tower blending.

Three-dimensional (3D) light-field display has achieved a great improvement. However, the collection of dense viewpoints in the real 3D scene is still a bottleneck. Virtual views can be generated by unsupervised networks, but the quality of different views is inconsistent because networks are separately trained on each posed view. Here, a virtual view synthesis method for the 3D light-field display based on scene tower blending is presented, which can synthesize high quality virtual views with correct occlusions by blending all tower results, and dense viewpoints on 3D light-field display can be provided with smooth motion parallax. Posed views are combinatorially input into diverse unsupervised CNNs to predict respective input-view towers, and towers of the same viewpoint are fused together. All posed-view towers are blended as a scene color tower and a scene selection tower, so that 3D scene distributions at different depth planes can be accurately estimated. Blended scene towers are soft-projected to synthesize virtual views with correct occlusions. A denoising network is used to improve the image quality of final synthetic views. Experimental results demonstrate the validity of the proposed method, which shows outstanding performances under various disparities. PSNR of the virtual views are about 30 dB and SSIM is above 0.91. We believe that our view synthesis method will be helpful for future applications of the 3D light-field display.

Disocclusion-type aware hole filling method for view synthesis

Depth-image-based rendering is an efficient way to produce content for 3D video and free viewpoint video. However, since the background that is occluded by the foreground objects in the reference view may become visible in the synthesized view, disocclusions are produced. In this paper, a disocclusion-type aware hole filling method is proposed for disocclusion handling. Disocclusions are divided into two types based on the depth value of their boundary pixels: foreground-background (FG-BG) disocclusion and background-background (BG-BG) disocclusion. For FG-BG disocclusion, the depth values of the associated pixels are optimized in the reference image to ensure the removal of ghosts and adaptively divide the disocclusion into some small holes. For BG-BG disocclusion, a foreground removal method is applied to remove the corresponding foreground objects. The removed regions are filled with the surrounding background textures so that the BG-BG disocclusion in the synthesized image can be eliminated. Experimental results indicate that the proposed method outperforms the other methods in the objective and subjective evaluations.

Real-Time Depth Video-Based Rendering for 6-DoF HMD Navigation and Light Field Displays

This paper presents a novel approach to provide immersive free navigation with 6 Degrees of Freedom in real-time for natural and virtual scenery, for both static and dynamic content. Stemming from the state-of-the-art in Depth Image-Based Rendering and the OpenGL pipeline, this new View Synthesis method achieves free navigation at up to 90 FPS and can take any number of input views with their corresponding depth maps as priors. Video content can be played thanks to GPU decompression, supporting free navigation with full parallax in real-time. To render a novel viewpoint, each selected input view is warped using the camera pose and associated depth map, using an implicit 3D representation. The warped views are then blended all together to generate the chosen virtual view. Various view blending approaches specifically designed to avoid visual artifacts are compared. Using as few as four input views appears to be an optimal trade-off between computation time and quality, allowing to synthesize high-quality stereoscopic views in real-time, offering a genuine immersive virtual reality experience. Additionally, the proposed approach provides high-quality rendering of a 3D scenery on holographic light field displays. Our results are comparable - objectively and subjectively - to the state of the art view synthesis tools NeRF and LLFF, while maintaining an overall lower complexity and real-time rendering.

High-quality holographic stereogram generation using four RGBD images.

To computer-generate high-quality holographic stereograms, a huge number of images must be provided: several hundred for a horizontal parallax and the square of this number for a full parallax. In this paper, we propose to drastically reduce this number to four input images with depth maps (or equivalently, four groups of neighboring images used to compute a depth map) in any pose, in order to create the missing images with depth image-based rendering. We evaluate the view synthesis method objectively before providing visual results of the corresponding holographic stereograms. We believe this method outperforms shearlet-based approaches in objective view synthesis quality metrics and in the number of required input images (7×7).

DIBR‐synthesised video quality assessment by measuring geometric distortion and spatiotemporal inconsistency

Depth-image-based rendering (DIBR), as the most popular view synthesis method, is commonly used in the application of multi-view and free-viewpoint videos. However, the quality evaluation of DIBR-synthesised videos remains largely unexplored, which may hinder the development of more advanced view synthesis technology. With this motivation, this Letter presents a new quality metric for DIBR-synthesised videos. Specifically, the disoccluded regions are first detected based on an adaptive threshold to quantify geometric distortions. An energy-based sequence mapping strategy is proposed to portray spatiotemporal inconsistency by calculating first-order and second-order similarities in the gradient magnitude domain and the Laplace-of-Gaussian domain, respectively. Finally, the overall quality score is generated by pooling the scores of geometric distortion and spatiotemporal inconsistency. Experimental results demonstrate that the proposed metric outperforms the state-of-the-art metrics dedicated to DIBR-synthesised images and videos.

Fast Hole Filling for View Synthesis in Free Viewpoint Video

The recent emergence of three-dimensional (3D) movies and 3D television (TV) indicates an increasing interest in 3D content. Stereoscopic displays have enabled visual experiences to be enhanced, allowing the world to be viewed in 3D. Virtual view synthesis is the key technology to present 3D content, and depth image-based rendering (DIBR) is a classic virtual view synthesis method. With a texture image and its corresponding depth map, a virtual view can be generated using the DIBR technique. The depth and camera parameters are used to project the entire pixel in the image to the 3D world coordinate system. The results in the world coordinates are then reprojected into the virtual view, based on 3D warping. However, these projections will result in cracks (holes). Hence, we herein propose a new method of DIBR for free viewpoint videos to solve the hole problem due to these projection processes. First, the depth map is preprocessed to reduce the number of holes, which does not produce large-scale geometric distortions; subsequently, improved 3D warping projection is performed collectively to create the virtual view. A median filter is used to filter the hole regions in the virtual view, followed by 3D inverse warping blending to remove the holes. Next, brightness adjustment and adaptive image blending are performed. Finally, the synthesized virtual view is obtained using the inpainting method. Experimental results verify that our proposed method can produce a pleasant visibility of the synthetized virtual view, maintain a high peak signal-to-noise ratio (PSNR) value, and efficiently decrease execution time compared with state-of-the-art methods.