Single-view 3D Reconstruction Research Articles

Primitive-Based 3D Human-Object Interaction Modelling and Programming

Embedding Human and Articulated Object Interaction (HAOI) in 3D is an important direction for a deeper human activity understanding. Different from previous works that use parametric and CAD models to represent humans and objects, in this work, we propose a novel 3D geometric primitive-based language to encode both humans and objects. Given our new paradigm, humans and objects are all compositions of primitives instead of heterogeneous entities. Thus, mutual information learning may be achieved between the limited 3D data of humans and different object categories. Moreover, considering the simplicity of the expression and the richness of the information it contains, we choose the superquadric as the primitive representation. To explore an effective embedding of HAOI for the machine, we build a new benchmark on 3D HAOI consisting of primitives together with their images and propose a task requiring machines to recover 3D HAOI using primitives from images. Moreover, we propose a baseline of single-view 3D reconstruction on HAOI. We believe this primitive-based 3D HAOI representation would pave the way for 3D HAOI studies. Our code and data are available at https://mvig-rhos.com/p3haoi.

Single-View 3D Reconstruction via Differentiable Rendering and Inverse Procedural Modeling

Three-dimensional models, reconstructed from real-life objects, are extensively used in virtual and mixed reality technologies. In this paper we propose an approach to 3D model reconstruction via inverse procedural modeling and describe two variants of this approach. The first option is to fit a set of input parameters using a genetic algorithm. The second option allows us to significantly improve precision by using gradients within the memetic algorithm, differentiable rendering, and differentiable procedural generators. We demonstrate the results of our work on different models, including trees, which are complex objects that most existing methods cannot reconstruct. In our work, we see two main contributions. First, we propose a method to join differentiable rendering and inverse procedural modeling. This gives us the ability to reconstruct 3D models more accurately than existing approaches when few input images are available, even for a single image. Second, we combine both differentiable and non-differentiable procedural generators into a single framework that allows us to apply inverse procedural modeling to fairly complex generators. We show that both variants of our approach can be useful: the differentiable one is more precise but puts limitations on the procedural generator, while the one based on genetic algorithms can be used with any existing generator. The proposed approach uses information about the symmetry and structure of the object to achieve high-quality reconstruction from a single image.

DeepMesh: Differentiable Iso-Surface Extraction.

Geometric Deep Learning has recently made striking progress with the advent of continuous deep implicit fields. They allow for detailed modeling of watertight surfaces of arbitrary topology while not relying on a 3D Euclidean grid, resulting in a learnable parameterization that is unlimited in resolution. Unfortunately, these methods are often unsuitable for applications that require an explicit mesh-based surface representation because converting an implicit field to such a representation relies on the Marching Cubes algorithm, which cannot be differentiated with respect to the underlying implicit field. In this work, we remove this limitation and introduce a differentiable way to produce explicit surface mesh representations from Deep Implicit Fields. Our key insight is that by reasoning on how implicit field perturbations impact local surface geometry, one can ultimately differentiate the 3D location of surface samples with respect to the underlying deep implicit field. We exploit this to define DeepMesh - an end-to-end differentiable mesh representation that can vary its topology. We validate our theoretical insight through several applications: Single view 3D Reconstruction via Differentiable Rendering, Physically-Driven Shape Optimization, Full Scene 3D Reconstruction from Scans and End-to-End Training. In all cases our end-to-end differentiable parameterization gives us an edge over state-of-the-art algorithms.

CMFAN: Cross-Modal Feature Alignment Network for Few-Shot Single-View 3D Reconstruction.

Few-shot single-view 3D reconstruction learns to reconstruct the novel category objects based on a query image and a few support shapes. However, since the query image and the support shapes are of different modalities, there is an inherent feature misalignment problem damaging the reconstruction. Previous works in the literature do not consider this problem. To this end, we propose the cross-modal feature alignment network (CMFAN) with two novel techniques. One is a strategy for model pretraining, namely, cross-modal contrastive learning (CMCL), here the 2D images and 3D shapes of the same objects compose the positives, and those from different objects form the negatives. With CMCL, the model learns to embed the 2D and 3D modalities of the same object into a tight area in the feature space and push away those from different objects, thus effectively aligning the global cross-modal features. The other is cross-modal feature fusion (CMFF), which further aligns and fuses the local features. Specifically, it first re-represents the local features with the cross-attention operation, making the local features share more information. Then, CMFF generates a descriptor for the support features and attaches it to each local feature vector of the query image with dense concatenation. Moreover, CMFF can be applied to multilevel local features and brings further advantages. We conduct extensive experiments to evaluate the effectiveness of our designs, and CMFAN sets new state-of-the-art performance in all of the 1-/10-/25-shot tasks of ShapeNet and ModelNet datasets.

FaceScape: 3D Facial Dataset and Benchmark for Single-View 3D Face Reconstruction.

In this paper, we present a large-scale detailed 3D face dataset, FaceScape, and the corresponding benchmark to evaluate single-view facial 3D reconstruction. By training on FaceScape data, a novel algorithm is proposed to predict elaborate riggable 3D face models from a single image input. FaceScape dataset releases 16,940 textured 3D faces, captured from 847 subjects and each with 20 specific expressions. The 3D models contain the pore-level facial geometry that is also processed to be topologically uniform. These fine 3D facial models can be represented as a 3D morphable model for coarse shapes and displacement maps for detailed geometry. Taking advantage of the large-scale and high-accuracy dataset, a novel algorithm is further proposed to learn the expression-specific dynamic details using a deep neural network. The learned relationship serves as the foundation of our 3D face prediction system from a single image input. Different from most previous methods, our predicted 3D models are riggable with highly detailed geometry under different expressions. We also use FaceScape data to generate the in-the-wild and in-the-lab benchmark to evaluate recent methods of single-view face reconstruction. The accuracy is reported and analyzed on the dimensions of camera pose and focal length, which provides a faithful and comprehensive evaluation and reveals new challenges. The unprecedented dataset, benchmark, and code have been released to the public for research purpose.

STAR-TM: STructure Aware Reconstruction of Textured Mesh From Single Image.

We present a novel method for single-view 3D reconstruction of textured meshes, with a focus to address the primary challenge surrounding texture inference and transfer. Our key observation is that learning textured reconstruction in a structure-aware and globally consistent manner is effective in handling the severe ill-posedness of the texturing problem and significant variations in object pose and texture details. Specifically, we perform structured mesh reconstruction, via a retrieval-and-assembly approach, to produce a set of genus-zero parts parameterized by deformable boxes and endowed with semantic information. For texturing, we first transfer visible colors from the input image onto the unified UV texture space of the deformable boxes. Then we combine a learned transformer model for per-part texture completion with a global consistency loss to optimize inter-part texture consistency. Our texture completion model operates in a VQ-VAE embedding space and is trained end-to-end, with the transformer training enhanced with retrieved texture instances to improve texture completion performance amid significant occlusion. Extensive experiments demonstrate higher-quality textured mesh reconstruction obtained by our method over state-of-the-art alternatives, both quantitatively and qualitatively, as reflected by a better recovery of texture coherence and details.

ISHS-Net: Single-View 3D Reconstruction by Fusing Features of Image and Shape Hierarchical Structures

The reconstruction of 3D shapes from a single view has been a longstanding challenge. Previous methods have primarily focused on learning either geometric features that depict overall shape contours but are insufficient for occluded regions, local features that capture details but cannot represent the complete structure, or structural features that encode part relationships but require predefined semantics. However, the fusion of geometric, local, and structural features has been lacking, leading to inaccurate reconstruction of shapes with occlusions or novel compositions. To address this issue, we propose a two-stage approach for achieving 3D shape reconstruction. In the first stage, we encode the hierarchical structure features of the 3D shape using an encoder-decoder network. In the second stage, we enhance the hierarchical structure features by fusing them with global and point features and feed the enhanced features into a signed distance function (SDF) prediction network to obtain rough SDF values. Using the camera pose, we project arbitrary 3D points in space onto different depth feature maps of the CNN and obtain their corresponding positions. Then, we concatenate the features of these corresponding positions together to form local features. These local features are also fed into the SDF prediction network to obtain fine-grained SDF values. By fusing the two sets of SDF values, we improve the accuracy of the model and enable it to reconstruct other object types with higher quality. Comparative experiments demonstrate that the proposed method outperforms state-of-the-art approaches in terms of accuracy.

Inferring implicit 3D representations from human figures on pictorial maps

ABSTRACT In this work, we present an automated workflow to bring human figures, one of the most frequently appearing entities on pictorial maps, to the third dimension. Our workflow is based on training data and neural networks for single-view 3D reconstruction of real humans from photos. We first let a network consisting of fully connected layers estimate the depth coordinate of 2D pose points. The gained 3D pose points are inputted together with 2D masks of body parts into a deep implicit surface network to infer 3D signed distance fields (SDFs). By assembling all body parts, we derive 2D depth images and body part masks of the whole figure for different views, which are fed into a fully convolutional network to predict UV images. These UV images and the texture for the given perspective are inserted into a generative network to inpaint the textures for the other views. The textures are enhanced by a cartoonization network and facial details are resynthesized by an autoencoder. Finally, the generated textures are assigned to the inferred body parts in a ray marcher. We test our workflow with 12 pictorial human figures after having validated several network configurations. The created 3D models look generally promising, especially when considering the challenges of silhouette-based 3D recovery and real-time rendering of the implicit SDFs. Further improvement is needed to reduce gaps between the body parts and to add pictorial details to the textures. Overall, the constructed figures may be used for animation and storytelling in digital 3D maps.

PA-Net: Plane Attention Network for real-time urban scene reconstruction

Traditional urban reconstruction methods can only output incomplete 3D models, which depict the scene regions that are visible to the moving camera. While learning-based shape reconstruction techniques make single-view 3D reconstruction possible, they are designed to handle single objects that are well-presented in the training datasets. This paper presents a novel learning-based approach for reconstructing complete 3D meshes for large-scale urban scenes in real-time. The input video sequences are fed into a localization module, which segments different objects and determines their relative positions. Each object is then reconstructed under their local coordinates to better approximate models in the training datasets. The reconstruction module is adopted from BSP-Net (Chen et al., 2020), which is capable of producing compact polygon meshes. However, major changes have been made so that unoriented objects in large-scale scenes can be reconstructed efficiently using only a small number of planes. Experimental results demonstrate that our approach can reconstruct urban scenes with buildings and vehicles using 400∼800 convex parts in 0.1∼0.5 s.

ED2IF2-Net: Learning Disentangled Deformed Implicit Fields and Enhanced Displacement Fields from Single Images Using Pyramid Vision Transformer

There has emerged substantial research in addressing single-view 3D reconstruction and the majority of the state-of-the-art implicit methods employ CNNs as the backbone network. On the other hand, transformers have shown remarkable performance in many vision tasks. However, it is still unknown whether transformers are suitable for single-view implicit 3D reconstruction. In this paper, we propose the first end-to-end single-view 3D reconstruction network based on the Pyramid Vision Transformer (PVT), called ED2IF2-Net, which disentangles the reconstruction of an implicit field into the reconstruction of topological structures and the recovery of surface details to achieve high-fidelity shape reconstruction. ED2IF2-Net uses a Pyramid Vision Transformer encoder to extract multi-scale hierarchical local features and a global vector of the input single image, which are fed into three separate decoders. A coarse shape decoder reconstructs a coarse implicit field based on the global vector, a deformation decoder iteratively refines the coarse implicit field using the pixel-aligned local features to obtain a deformed implicit field through multiple implicit field deformation blocks (IFDBs), and a surface detail decoder predicts an enhanced displacement field using the local features with hybrid attention modules (HAMs). The final output is a fusion of the deformed implicit field and the enhanced displacement field, with four loss terms applied to reconstruct the coarse implicit field, structure details through a novel deformation loss, overall shape after fusion, and surface details via a Laplacian loss. The quantitative results obtained from the ShapeNet dataset validate the exceptional performance of ED2IF2-Net. Notably, ED2IF2-Net-L stands out as the top-performing variant, exhibiting the highest mean IoU, CD, EMD, ECD-3D, and ECD-2D scores, reaching impressive values of 61.1, 7.26, 2.51, 6.08, and 1.84, respectively. The extensive experimental evaluations consistently demonstrate the state-of-the-art capabilities of ED2IF2-Net in terms of reconstructing topological structures and recovering surface details, all while maintaining competitive inference time.

RAFaRe: Learning Robust and Accurate Non-parametric 3D Face Reconstruction from Pseudo 2D&3D Pairs

We propose a robust and accurate non-parametric method for single-view 3D face reconstruction (SVFR). While tremendous efforts have been devoted to parametric SVFR, a visible gap still lies between the result 3D shape and the ground truth. We believe there are two major obstacles: 1) the representation of the parametric model is limited to a certain face database; 2) 2D images and 3D shapes in the fitted datasets are distinctly misaligned. To resolve these issues, a large-scale pseudo 2D&3D dataset is created by first rendering the detailed 3D faces, then swapping the face in the wild images with the rendered face. These pseudo 2D&3D pairs are created from publicly available datasets which eliminate the gaps between 2D and 3D data while covering diverse appearances, poses, scenes, and illumination. We further propose a non-parametric scheme to learn a well-generalized SVFR model from the created dataset, and the proposed hierarchical signed distance function turns out to be effective in predicting middle-scale and small-scale 3D facial geometry. Our model outperforms previous methods on FaceScape-wild/lab and MICC benchmarks and is well generalized to various appearances, poses, expressions, and in-the-wild environments. The code is released at https://github.com/zhuhao-nju/rafare.

Machine learning based geometry reconstruction for quality control of laser welding processes

Abstract The increasing use of automated laser welding processes causes high demands on quality control. 2D or 3D sensor technology can be used for data acquisition to monitor the weld quality after laser welding. Compared to a 2D camera image, the 3D height data, e.g. acquired using optical coherence tomography, contains additional relevant information for quality inspection. However, the disadvantages are system complexity, higher costs, and longer acquisition times. Therefore, we compare image-based methods with the quality assessment based on height data. The first method uses feature vectors from grayscale images taken coaxially with the laser beam. The significant advantage is that a camera is often integrated into the laser system, so no additional hardware is required. In the second approach, we use an AI-based single-view 3D reconstruction method. The height profile is reconstructed from a camera image and used for further quality assessment. Thus, we combine the advantages of 2D data acquisition with higher accuracy in evaluating 3D data. In addition, we consider the usually low data availability in the industrial environment in the development of algorithms. We use a training data set with 95 samples and a test data set with 858 samples. The work uses the contracting process of copper wires to produce formed coil windings to illustrate the method. We analyze a data set with different defect types and compare the quality assessment using the height data acquired with OCT, the feature vectors from the camera images, and the reconstructed height data.

Adaptive Points Sampling for Implicit Field Reconstruction of Industrial Digital Twin.

Nowadays, the digital twin (DT) plays an important role in Industry 4.0. It aims to model reality in the digital space for further industrial maintenance, management, and optimization. Previously, many AI technologies have been applied in this field and provide strong tools to connect physical and virtual spaces. However, we found that single-view 3D reconstruction (SVR) for DT has not been thoroughly studied. SVR can generate 3D digital models of real industrial products from just a single image. The application of SVR technology would bring convenience, cheapness, and robustness to modeling physical objects in digital space. However, the existing SVR methods cannot perform well in the reconstruction of details, which is indispensable and challenging in industrial products. In this paper, we propose a new detail-aware feature extraction network based on a feature pyramid network (FPN) for better detail reconstruction. Then, an extra network is designed to combine convolutional feature maps from different levels. Moreover, we also propose a novel adaptive points-sampling strategy to adaptively change the learning difficulty according to the training status. This can accelerate the training process and improve the fine-tuned network performance as well. Finally, we conduct comprehensive experiments on both the general objects dataset ShapeNet and a collected industrial dataset to prove the effectiveness of our methods and the practicability of the SVR technology for DT.

3D‐FEGNet: A feature enhanced point cloud generation network from a single image

Deep learning-based single view 3D reconstruction is a hot topic in computer vision. However, predicting a more realistic 3D point cloud from a single image is an ill-posed problem. In recent years, most of the 3D point cloud prediction methods based on single view are straight-through structure, which will cause the loss of part of feature information and the loss of part of detail information of the resulting point clouds, which will lead to the unsatisfactory visual effect of reconstructed point clouds. In this paper, a Feature-Enhanced 3D point clouds generation Network (3D-FENet) from a single image is proposed. In order to enhance the feature information of RGB image, edge extraction module is adopted. In the process of point cloud generation, a point cloud pyramid is designed, which combines low resolution point cloud with high resolution point cloud to enhance the local details of the generated point clouds. In the fine-tuning stage, the differential projection module is used to fine-tune the whole network by 2D projection of reconstructed point clouds. Experimental results show that the performance of the authors’ proposed method is better than the state-of-the-art studies.

Analysis of AI-Based Single-View 3D Reconstruction Methods for an Industrial Application.

Machine learning (ML) is a key technology in smart manufacturing as it provides insights into complex processes without requiring deep domain expertise. This work deals with deep learning algorithms to determine a 3D reconstruction from a single 2D grayscale image. The potential of 3D reconstruction can be used for quality control because the height values contain relevant information that is not visible in 2D data. Instead of 3D scans, estimated depth maps based on a 2D input image can be used with the advantage of a simple setup and a short recording time. Determining a 3D reconstruction from a single input image is a difficult task for which many algorithms and methods have been proposed in the past decades. In this work, three deep learning methods, namely stacked autoencoder (SAE), generative adversarial networks (GANs) and U-Nets are investigated, evaluated and compared for 3D reconstruction from a 2D grayscale image of laser-welded components. In this work, different variants of GANs are tested, with the conclusion that Wasserstein GANs (WGANs) are the most robust approach among them. To the best of our knowledge, the present paper considers for the first time the U-Net, which achieves outstanding results in semantic segmentation, in the context of 3D reconstruction tasks. Unlike the U-Net, which uses standard convolutions, the stacked dilated U-Net (SDU-Net) applies stacked dilated convolutions. Of all the 3D reconstruction approaches considered in this work, the SDU-Net shows the best performance, not only in terms of evaluation metrics but also in terms of computation time. Due to the comparably small number of trainable parameters and the suitability of the architecture for strong data augmentation, a robust model can be generated with only a few training data.

A single view leaf reconstruction method based on the fusion of ResNet and differentiable render in plant growth digital twin system

• A single view leaf reconstruction based on deep learning for plant growth digital twin system is proposed. • The model provides a feature extraction network with RestNet18 to improve pixel2mesh’ feature extraction capabilities. • The model provides a way to optimize the shape of the leaf mesh with differentiable rendering. In modern agriculture, plant growth digital twin system helps breeders monitor plant growth, increase yield, and provide growth management advice. Research on the single view leaf 3D reconstruction in digital twin systems has achieved relative success. However, in traditional single-view reconstruction algorithms, the leaf reconstruction often contains the problems of low precision, achieving complexity, and slow speed, making it difficult for recovering three-dimensional information about leaves. Consequently, the reconstruction precision is significantly reduced, which further affects the accuracy of single-view leaf 3D reconstruction. In response to this problem, this study proposed a single-view leaf reconstruction approach in plant growth digital twin systems based on deep learning. The method in this paper mainly fuses the advantages of ResNet and differentiable rendering, and the model is used for further enhancing feature extraction capability and reconstruction precision. Finally, the experiment presented in this paper suggests that the method allows for the 3D reconstruction of plant leaves with different shapes using a single view. Moreover, the experiment results show that the F-Score, CD, EMD reached 76.192, 0.808, and 3.567. Compared with other models, the proposed model in this study has higher reconstruction accuracy, 3D evaluation indicators, and prediction results, providing important ideas and methods for recovering the leaves from a single view in a plant growth digital twin system.

Sym3DNet: Symmetric 3D Prior Network for Single-View 3D Reconstruction.

The three-dimensional (3D) symmetry shape plays a critical role in the reconstruction and recognition of 3D objects under occlusion or partial viewpoint observation. Symmetry structure prior is particularly useful in recovering missing or unseen parts of an object. In this work, we propose Sym3DNet for single-view 3D reconstruction, which employs a three-dimensional reflection symmetry structure prior of an object. More specifically, Sym3DNet includes 2D-to-3D encoder-decoder networks followed by a symmetry fusion step and multi-level perceptual loss. The symmetry fusion step builds flipped and overlapped 3D shapes that are fed to a 3D shape encoder to calculate the multi-level perceptual loss. Perceptual loss calculated in different feature spaces counts on not only voxel-wise shape symmetry but also on the overall global symmetry shape of an object. Experimental evaluations are conducted on both large-scale synthetic 3D data (ShapeNet) and real-world 3D data (Pix3D). The proposed method outperforms state-of-the-art approaches in terms of efficiency and accuracy on both synthetic and real-world datasets. To demonstrate the generalization ability of our approach, we conduct an experiment with unseen category samples of ShapeNet, exhibiting promising reconstruction results as well.

Single-View 3d Reconstruction of Surface of Revolution

Single-View 3d Reconstruction of Surface of Revolution

Single-View 3D reconstruction: A Survey of deep learning methods

The field of single-view 3D shape reconstruction and generation using deep learning techniques has seen rapid growth in the past five years. As the field is reaching a stage of maturity, a plethora of methods has been continuously proposed with the aim of pushing the state of research further. This article focuses on surveying the literature by classifying these methods according to the shape representation they use as an output. Specifically, it covers each method’s main contributions, degree of supervision, training paradigm, and its relation to the whole body of literature. Additionally, this survey discusses common 3D datasets, loss functions, and evaluation metrics used in the field. Finally, it provides a thorough analysis and reflections on the current state of research and provides a summary of the open problems and possible future directions. This work is an effort to introduce the field of data-driven single-view 3D reconstruction to interested researchers while being comprehensive enough to act as a reference to those who already do research in the field.

Single View 3D Reconstruction Based on Improved RGB-D Image

Image-based 3D reconstruction is a fundamental task in computer vision and computer graphics. Although many existing approaches have obtained excellent results, they all dependent on multi-view images, then these methods are very complex. To reduce the complexity of the existing approaches, in this paper we fist propose a novel 3D reconstruction method based on single RGB image and the corresponding depth image, and design a unified framework called Single3D. Second, we trained a neural network model to reduce noises for improving the quality of the input RGB image, then leading a desirable RGB image. Third, we combine the deep guided filter and fast guided filter to fill holes on the depth image for improving the quality of 3D model. Fourth, we recover high-quality 3D model from the RGB-D image based on the proposed Single3D framework. Finally, the comprehensive experiments were conducted on the most challenging benchmarking dataset with variant light, occlusion, and clustering. Experimental results show that our method has desirable accuracy and efficiency.