Fine-grained Details Research Articles

Forecasting with an N-dimensional Langevin equation and a neural-ordinary differential equation.

Accurate prediction of electricity day-ahead prices is essential in competitive electricity markets. Although stationary electricity-price forecasting techniques have received considerable attention, research on non-stationary methods is comparatively scarce, despite the common prevalence of non-stationary features in electricity markets. Specifically, existing non-stationary techniques will often aim to address individual non-stationary features in isolation, leaving aside the exploration of concurrent multiple non-stationary effects. Our overarching objective here is the formulation of a framework to systematically model and forecast non-stationary electricity-price time series, encompassing the broader scope of non-stationary behavior. For this purpose, we develop a data-driven model that combines an N-dimensional Langevin equation (LE) with a neural-ordinary differential equation (NODE). The LE captures fine-grained details of the electricity-price behavior in stationary regimes but is inadequate for non-stationary conditions. To overcome this inherent limitation, we adopt a NODE approach to learn, and at the same time predict, the difference between the actual electricity-price time series and the simulated price trajectories generated by the LE. By learning this difference, the NODE reconstructs the non-stationary components of the time series that the LE is not able to capture. We exemplify the effectiveness of our framework using the Spanish electricity day-ahead market as a prototypical case study. Our findings reveal that the NODE nicely complements the LE, providing a comprehensive strategy to tackle both stationary and non-stationary electricity-price behavior. The framework's dependability and robustness is demonstrated through different non-stationary scenarios by comparing it against a range of basic naïve methods.

SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation

Recently, self-supervised monocular depth estimation has gained popularity with numerous applications in autonomous driving and robotics. However, existing solutions primarily seek to estimate depth from immediate visual features, and struggle to recover fine-grained scene details. In this paper, we introduce SQLdepth, a novel approach that can effectively learn fine-grained scene structure priors from ego-motion. In SQLdepth, we propose a novel Self Query Layer (SQL) to build a self-cost volume and infer depth from it, rather than inferring depth from feature maps. We show that, the self-cost volume is an effective inductive bias for geometry learning, which implicitly models the single-frame scene geometry, with each slice of it indicating a relative distance map between points and objects in a latent space. Experimental results on KITTI and Cityscapes show that our method attains remarkable state-of-the-art performance, and showcases computational efficiency, reduced training complexity, and the ability to recover fine-grained scene details. Moreover, the self-matching-oriented relative distance querying in SQL improves the robustness and zero-shot generalization capability of SQLdepth. Code is available at https://github.com/hisfog/SfMNeXt-Impl.

SC-NeuS: Consistent Neural Surface Reconstruction from Sparse and Noisy Views

The recent neural surface reconstruction approaches using volume rendering have made much progress by achieving impressive surface reconstruction quality, but are still limited to dense and highly accurate posed views. To overcome such drawbacks, this paper pays special attention on the consistent surface reconstruction from sparse views with noisy camera poses. Unlike previous approaches, the key difference of this paper is to exploit the multi-view constraints directly from the explicit geometry of the neural surface, which can be used as effective regularization to jointly learn the neural surface and refine the camera poses. To build effective multi-view constraints, we introduce a fast differentiable on-surface intersection to generate on-surface points, and propose view-consistent losses on such differentiable points to regularize the neural surface learning. Based on this point, we propose a joint learning strategy, named SC-NeuS, to perform geometry-consistent surface reconstruction in an end-to-end manner. With extensive evaluation on public datasets, our SC-NeuS can achieve consistently better surface reconstruction results with fine-grained details than previous approaches, especially from sparse and noisy camera views. The source code is available at https://github.com/zouzx/sc-neus.git.

Spatial-Contextual Discrepancy Information Compensation for GAN Inversion

Most existing GAN inversion methods either achieve accurate reconstruction but lack editability or offer strong editability at the cost of fidelity. Hence, how to balance the distortion-editability trade-off is a significant challenge for GAN inversion. To address this challenge, we introduce a novel spatial-contextual discrepancy information compensation-based GAN-inversion method (SDIC), which consists of a discrepancy information prediction network (DIPN) and a discrepancy information compensation network (DICN). SDIC follows a ``compensate-and-edit'' paradigm and successfully bridges the gap in image details between the original image and the reconstructed/edited image. On the one hand, DIPN encodes the multi-level spatial-contextual information of the original and initial reconstructed images and then predicts a spatial-contextual guided discrepancy map with two hourglass modules. In this way, a reliable discrepancy map that models the contextual relationship and captures fine-grained image details is learned. On the other hand, DICN incorporates the predicted discrepancy information into both the latent code and the GAN generator with different transformations, generating high-quality reconstructed/edited images. This effectively compensates for the loss of image details during GAN inversion. Both quantitative and qualitative experiments demonstrate that our proposed method achieves the excellent distortion-editability trade-off at a fast inference speed for both image inversion and editing tasks. Our code is available at https://github.com/ZzqLKED/SDIC.

PrefAce: Face-Centric Pretraining with Self-Structure Aware Distillation

Video-based facial analysis is important for autonomous agents to understand human expressions and sentiments. However, limited labeled data is available to learn effective facial representations. This paper proposes a novel self-supervised face-centric pretraining framework, called PrefAce, which learns transferable video facial representation without labels. The self-supervised learning is performed with an effective landmark-guided global-local tube distillation. Meanwhile, a novel instance-wise update FaceFeat Cache is built to enforce more discriminative and diverse representations for downstream tasks. Extensive experiments demonstrate that the proposed framework learns universal instance-aware facial representations with fine-grained landmark details from videos. The point is that it can transfer across various facial analysis tasks, e.g., Facial Attribute Recognition (FAR), Facial Expression Recognition (FER), DeepFake Detection (DFD), and Lip Synchronization (LS). Our framework also outperforms the state-of-the-art on various downstream tasks, even in low data regimes. Code is available at https://github.com/siyuan-h/PrefAce.

VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook

Night photography often struggles with challenges like low light and blurring, stemming from dark environments and prolonged exposures. Current methods either disregard priors and directly fitting end-to-end networks, leading to inconsistent illumination, or rely on unreliable handcrafted priors to constrain the network, thereby bringing the greater error to the final result. We believe in the strength of data-driven high-quality priors and strive to offer a reliable and consistent prior, circumventing the restrictions of manual priors. In this paper, we propose Clearer Night Image Restoration with Vector-Quantized Codebook (VQCNIR) to achieve remarkable and consistent restoration outcomes on real-world and synthetic benchmarks. To ensure the faithful restoration of details and illumination, we propose the incorporation of two essential modules: the Adaptive Illumination Enhancement Module (AIEM) and the Deformable Bi-directional Cross-Attention (DBCA) module. The AIEM leverages the inter-channel correlation of features to dynamically maintain illumination consistency between degraded features and high-quality codebook features. Meanwhile, the DBCA module effectively integrates texture and structural information through bi-directional cross-attention and deformable convolution, resulting in enhanced fine-grained detail and structural fidelity across parallel decoders. Extensive experiments validate the remarkable benefits of VQCNIR in enhancing image quality under low-light conditions, showcasing its state-of-the-art performance on both synthetic and real-world datasets. The code is available at https://github.com/AlexZou14/VQCNIR.

Terrain Diffusion Network: Climatic-Aware Terrain Generation with Geological Sketch Guidance

Sketch-based terrain generation seeks to create realistic landscapes for virtual environments in various applications such as computer games, animation and virtual reality. Recently, deep learning based terrain generation has emerged, notably the ones based on generative adversarial networks (GAN). However, these methods often struggle to fulfill the requirements of flexible user control and maintain generative diversity for realistic terrain. Therefore, we propose a novel diffusion-based method, namely terrain diffusion network (TDN), which actively incorporates user guidance for enhanced controllability, taking into account terrain features like rivers, ridges, basins, and peaks. Instead of adhering to a conventional monolithic denoising process, which often compromises the fidelity of terrain details or the alignment with user control, a multi-level denoising scheme is proposed to generate more realistic terrains by taking into account fine-grained details, particularly those related to climatic patterns influenced by erosion and tectonic activities. Specifically, three terrain synthesisers are designed for structural, intermediate, and fine-grained level denoising purposes, which allow each synthesiser concentrate on a distinct terrain aspect. Moreover, to maximise the efficiency of our TDN, we further introduce terrain and sketch latent spaces for the synthesizers with pre-trained terrain autoencoders. Comprehensive experiments on a new dataset constructed from NASA Topology Images clearly demonstrate the effectiveness of our proposed method, achieving the state-of-the-art performance. Our code is available at https://github.com/TDNResearch/TDN.

Multi-Modal Disordered Representation Learning Network for Description-Based Person Search

Description-based person search aims to retrieve images of the target identity via textual descriptions. One of the challenges for this task is to extract discriminative representation from images and descriptions. Most existing methods apply the part-based split method or external models to explore the fine-grained details of local features, which ignore the global relationship between partial information and cause network instability. To overcome these issues, we propose a Multi-modal Disordered Representation Learning Network (MDRL) for description-based person search to fully extract the visual and textual representations. Specifically, we design a Cross-modality Global Feature Learning Architecture to learn the global features from the two modalities and meet the demand of the task. Based on our global network, we introduce a Disorder Local Learning Module to explore local features by a disordered reorganization strategy from both visual and textual aspects and enhance the robustness of the whole network. Besides, we introduce a Cross-modality Interaction Module to guide the two streams to extract visual or textual representations considering the correlation between modalities. Extensive experiments are conducted on two public datasets, and the results show that our method outperforms the state-of-the-art methods on CUHK-PEDES and ICFG-PEDES datasets and achieves superior performance.

Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning

Recently, arbitrary-scale point cloud upsampling mechanism became increasingly popular due to its efficiency and convenience for practical applications. To achieve this, most previous approaches formulate it as a problem of surface approximation and employ point-based networks to learn surface representations. However, learning surfaces from sparse point clouds is more challenging, and thus they often suffer from the low-fidelity geometry approximation. To address it, we propose an arbitrary-scale Point cloud Upsampling framework using Voxel-based Network (PU-VoxelNet). Thanks to the completeness and regularity inherited from the voxel representation, voxel-based networks are capable of providing predefined grid space to approximate 3D surface, and an arbitrary number of points can be reconstructed according to the predicted density distribution within each grid cell. However, we investigate the inaccurate grid sampling caused by imprecise density predictions. To address this issue, a density-guided grid resampling method is developed to generate high-fidelity points while effectively avoiding sampling outliers. Further, to improve the fine-grained details, we present an auxiliary training supervision to enforce the latent geometric consistency among local surface patches. Extensive experiments indicate the proposed approach outperforms the state-of-the-art approaches not only in terms of fixed upsampling rates but also for arbitrary-scale upsampling. The code is available at https://github.com/hikvision-research/3DVision

A point contextual transformer network for point cloud completion

Point cloud completion is an essential task for recovering a complete point cloud from its partial observation to support downstream applications, such as object detection and reconstruction. Existing point cloud completion networks primarily rely on large-scale datasets to learn the mapping between the partial shapes and the complete shapes. They often adopt a multi-stage strategy to progressively generate complete point clouds with finer details. However, underutilization of shape priors and complex modelling frameworks still plague these networks. To address these issues, we innovatively propose a point contextual transformer (PCoT) for point cloud completion (PCoT-Net). We design the PCoT to adaptively fuse static and dynamic point contextual information. This allows for the effective capture of fine-grained local contextual features. We then propose a one-stage network with a feature completion module to directly generate credible and detailed complete point cloud results. Furthermore, we incorporate External Attention (EA) into the feature completion module, which is lightweight and further improves the performance of learning complete features and reconstructing the complete point cloud. Extensive experiments on various datasets validate the effectiveness of our PCoT-based approach and the EA-enhanced feature completion module, which achieves superior quantitative performance in Chamfer Distance (CD) and F1-Score. In comparison to PMP-Net++ (Wen et al., 2022), our method improves the F1-Score by 0.010, 0.022, and 0.026, and reduces the CD by 0.16, 0.95, and 1.74 on the MVP, CRN, and ScanNet datasets, respectively, while providing visually superior results, capturing more fine-grained details and producing smoother reconstructed surfaces.

A comprehensive research of deep learning approaches for High Definition map construction in autonomous driving

In the realm of autonomous driving, High Definition Maps (HD maps) are indispensable for safe and precise navigation. Traditional HD map construction methods involving point cloud capture and SLAM have proven effective but labor-intensive. This paper addresses the growing interest in leveraging deep learning techniques to streamline HD map creation. This paper presents a systematic exploration of deep learning methodologies for HD map construction. It categorizes these approaches into two core components: Feature Extraction and Feature Decoding. Feature Extraction involves the transformation of input data, comprising images and LiDAR point clouds, into Bird's Eye View (BEV) representations. Feature Decoding is dissected into rasterized map objectives and vector map objectives. Detailed analysis is conducted on prominent methodologies. The paper provides a nuanced evaluation of these deep learning techniques, highlighting their respective strengths and limitations. Factors such as precision, computational efficiency, and the preservation of fine-grained details are considered when selecting the most suitable method. This comprehensive review summarizes and prospects the research in related fields.

Mask2Former with Improved Query for Semantic Segmentation in Remote-Sensing Images

Semantic segmentation of remote sensing (RS) images is vital in various practical applications, including urban construction planning, natural disaster monitoring, and land resources investigation. However, RS images are captured by airplanes or satellites at high altitudes and long distances, resulting in ground objects of the same category being scattered in various corners of the image. Moreover, objects of different sizes appear simultaneously in RS images. For example, some objects occupy a large area in urban scenes, while others only have small regions. Technically, the above two universal situations pose significant challenges to the segmentation with a high quality for RS images. Based on these observations, this paper proposes a Mask2Former with an improved query (IQ2Former) for this task. The fundamental motivation behind the IQ2Former is to enhance the capability of the query of Mask2Former by exploiting the characteristics of RS images well. First, we propose the Query Scenario Module (QSM), which aims to learn and group the queries from feature maps, allowing the selection of distinct scenarios such as the urban and rural areas, building clusters, and parking lots. Second, we design the query position module (QPM), which is developed to assign the image position information to each query without increasing the number of parameters, thereby enhancing the model’s sensitivity to small targets in complex scenarios. Finally, we propose the query attention module (QAM), which is constructed to leverage the characteristics of query attention to extract valuable features from the preceding queries. Being positioned between the duplicated transformer decoder layers, QAM ensures the comprehensive utilization of the supervisory information and the exploitation of those fine-grained details. Architecturally, the QSM, QPM, and QAM as well as an end-to-end model are assembled to achieve high-quality semantic segmentation. In comparison to the classical or state-of-the-art models (FCN, PSPNet, DeepLabV3+, OCRNet, UPerNet, MaskFormer, Mask2Former), IQ2Former has demonstrated exceptional performance across three publicly challenging remote-sensing image datasets, 83.59 mIoU on the Vaihingen dataset, 87.89 mIoU on Potsdam dataset, and 56.31 mIoU on LoveDA dataset. Additionally, overall accuracy, ablation experiment, and visualization segmentation results all indicate IQ2Former validity.

PATNet: Patch-to-pixel attention-aware transformer network for RGB-D and RGB-T salient object detection

Multimodal salient object detection (SOD) combines different modal images to generate the most visually appealing saliency map. When fusing multimodal and multiscale features, maintaining the integrity and fine granularity of the target is critical for improving the performance of multimodal SOD. The fine-grained information differences between the modalities and the size of the features in the transformer prevent most existing studies from guaranteeing both granularities. Therefore, we propose a patch-to-pixel attention-aware transformer network (PATNet) to overcome these problems, whereby the integrity and fine-grained details of the saliency map are preserved by employing a decision-transformation strategy to map global patches onto local pixels. Specifically, PATNet consists of the shared attention fusion module (SAFM), adjacent modeling fusion module (AMFM), and fine-grained mapping module (FMM). SAFM enhances the consistency between multimodal features through a shared attention matrix and an identical convolutional feed-forward network. Meanwhile, AMFM enhances low-resolution features by modeling neighboring features to avoid the aliasing effect of upsampling. In the output stage, FMM is responsible for mapping the feature maps represented by patches onto pixels and restoring the salient object details. Numerous experimental results demonstrate that PATNet outperforms 24 state-of-the-art methods on six RGB-D and three RGB-T datasets. The source code is publicly available at https://github.com/LitterMa-820/PATNet.

A Practical Deep Learning Architecture for Large-Area Solid Wastes Monitoring Based on UAV Imagery

The development of global urbanization has brought about a significant amount of solid waste. These untreated wastes may be dumped in any corner, causing serious pollution to the environment. Thus, it is necessary to accurately obtain their distribution locations and detailed edge information. In this study, a practical deep learning network for recognizing solid waste piles over extensive areas using unmanned aerial vehicle (UAV) imagery has been proposed and verified. Firstly, a high-resolution dataset serving to solid waste detection was created based on UAV aerial data. Then, a dual-branch solid waste semantic segmentation model was constructed to address the characteristics of the integration of solid waste distribution with the environment and the irregular edge morphology. The Context feature branch is responsible for extracting high-level semantic features, while the Spatial feature branch is designed to capture fine-grained spatial details. After information fusion, the model obtained more comprehensive feature representation and segmentation ability. The effectiveness of the improvement was verified through ablation experiments and compared with 13 commonly used semantic segmentation models, demonstrating the advantages of the method in solid waste segmentation tasks, with an overall accuracy of over 94%, and a recall rate of 88.6%—much better than the best performing baselines. Finally, a spatial distribution map of solid waste over Jiaxing district, China was generated by the model inference, which assisted the environmental protection department in completing environmental management. The proposed method provides a feasible approach for the accurately monitoring of solid waste, so as to provide policy support for environmental protection.

Hierarchical matching and reasoning for multi-query image retrieval

As a promising field, Multi-Query Image Retrieval (MQIR) aims at searching for the semantically relevant image given multiple region-specific text queries. Existing works mainly focus on a single-level similarity between image regions and text queries, which neglect the hierarchical guidance of multi-level similarities and result in incomplete alignments. Besides, the high-level semantic correlations that intrinsically connect different region–query pairs are rarely considered. To address above limitations, we propose a novel Hierarchical Matching and Reasoning Network (HMRN) for MQIR. It disentangles MQIR into three hierarchical semantic representations, which is responsible to capture fine-grained local details, contextual global scopes, and high-level inherent correlations. HMRN consists of two modules: Scalar-based Matching (SM) module and Vector-based Reasoning (VR) module. Specifically, the SM module characterizes the multi-level alignment similarity, which consists of a fine-grained local-level similarity and a context-aware global-level similarity. Afterwards, the VR module is developed to excavate the potential semantic correlations among multiple region–query pairs, which further explores the high-level reasoning similarity. Finally, these three-level similarities are aggregated into a joint similarity space to form the ultimate similarity. Extensive experiments on the benchmark dataset demonstrate that our HMRN substantially surpasses the current state-of-the-art methods. For instance, compared with the existing best method Drill-down, the metric R@1 in the last round is improved by 23.4%. Our source codes will be released at https://github.com/LZH-053/HMRN.

Multi-depth branch network for efficient image super-resolution

A longstanding challenge in Super-Resolution (SR) is how to efficiently enhance high-frequency details in Low-Resolution (LR) images while maintaining semantic coherence. This is particularly crucial in practical applications where SR models are often deployed on low-power devices. To address this issue, we propose an innovative asymmetric SR architecture featuring Multi-Depth Branch Module (MDBM). These MDBMs contain branches of different depths, designed to capture high- and low-frequency information simultaneously and efficiently. The hierarchical structure of MDBM allows the deeper branch to gradually accumulate fine-grained local details under the contextual guidance of the shallower branch. We visualize this process using feature maps, and further demonstrate the rationality and effectiveness of this design using proposed novel Fourier spectral analysis methods. Moreover, our model exhibits more significant spectral differentiation between branches than existing branch networks. This suggests that MDBM reduces feature redundancy and offers a more effective method for integrating high- and low-frequency information. Extensive qualitative and quantitative evaluations on various datasets show that our model can generate structurally consistent and visually realistic HR images. It achieves state-of-the-art (SOTA) results at a very fast inference speed. Our code is available at https://github.com/thy960112/MDBN.

MULTI-LABEL CLASSIFICATION OF THORACIC DISEASES USING STRUCTURED DEEP LEARNING FRAMEWORK

The current medical scenario indicates that thoracic diseases are the primary cause of illnesses of human beings worldwide. COVID-19 outbreak results in the generation of tremendous Chest X-Ray (CXR) and Computed Tomography (CT) image data archives because CXR and CT imaging are the foremost effective screening tools that characterize patterns and features of pathologies present in the lungs. After careful observation, these chest radiographs show that images are often labeled with more than one pathology, that extends disease diagnosis problem to tedious multi-label classification task. So, this paper proposes a modified mish activation function and compound loss-based Separable Convolution Neural Network (SCNN) model that accomplishes image-level detection of multiple pathologies and COVID-19 from multi-modality chest radiographs. We investigate the power of separable convolution in the CNN network that interprets spatial and depth-wise dimensions of each pixel in CXR and CT images. A modified mish called as Swmish, a non-monotonic activation function, is introduced that preserves negative gradient flow of input and capture fine-grained detail of multiple regions. Moreover, a compound loss function of dice coefficient and binary cross entropy is applied that strongly optimizes the model during training. The generality of the proposed SCNN model is confirmed after conducting comprehensive experiments on multi-label CXR and COVID19-CT datasets. The proposed model gains an AUC score of 0.91 and loss of 0.13 on the multi-label CXR dataset and achieves a testing accuracy of 0.97, AUC score of 0.99 and a loss of 0.11 on the COVID19-CT dataset. These results show that the proposed model achieves highest evaluation metrics in classifying binary and multi-label thoracic abnormalities from multi-modality chest radiographs.

DBNet-SI: Dual branch network of shift window attention and inception structure for skin lesion segmentation

The U-shaped convolutional neural network (CNN) has attained remarkable achievements in the segmentation of skin lesion. However, given the inherent locality of convolution, this architecture cannot capture long-range pixel dependencies and multiscale global contextual information effectively. Moreover, repeated convolutions and downsampling operations can readily result in the omission of intricate local fine-grained details. In this paper, we proposed a U-shaped network (DBNet-SI) equipped with a dual-branch module that combines shift window attention and inception structures. First, we proposed a dual-branch module that combines shift window attention and inception structures (MSI) to better capture multiscale global contextual information and long-range pixel dependencies. Specifically, we have devised a cross-branch bidirectional interaction module within the MSI module to enable information complementarity between the two branches in the channel and spatial dimensions. Therefore, MSI is capable of extracting distinguishing and comprehensive features to accurately identify the skin lesion boundaries. Second, we have devised a progressive feature enhancement and information compensation module (PFEIC), which progressively compensates for fine-grained features through reconstructed skip connections and integrated global context attention modules. The results of the experiment show the superior segmentation performance of DBNet-SI compared with other deep learning models for skin lesion segmentation in the ISIC2017 and ISIC2018 datasets. Ablation studies demonstrate that our model can effectively extract rich multiscale global contextual information and compensate for the loss of local details.

An Intelligent Traffic Analysis and Prediction System Using Deep Learning Technique

Accurate identification of vehicles and estimating density in traffic surveillance systems is a challenging task, particularly in scenarios with closely spaced lanes. Single Shot MultiBox Detector (SSD) is introduced in vehicle detection and classification due to its speed and accuracy. It utilizes a transfer learning technique that enables them to utilize features from pretrained Convolutional Neural Networks (CNNs). Although the utilization of multi-scale feature maps in SSD has achieved efficient results in vehicle detection, it struggles to identify key features of vehicles due to its one-stage detection approach. Furthermore, the use of VGG16 as the backbone network in these approaches leads to the loss of fine-grained details, posing challenges in accurately localizing and classifying small vehicles. Also, there is no interaction between high- and low-level features, which restricts the networks ability to effectively integrate and utilize both features for accurate vehicle detection. To overcome these limitations, an improved SSD approach is introduced in this paper, leveraging interactive multi-scale attention characteristics to accurately detect and classify vehicles. This approach utilizes ResNet50 as the network backbone to overcome the limitations of traditional SSDs in detecting small vehicles. Also, an attention block is incorporated into it to focus on key details and assign higher attention to relevant pixels within the feature map. Also, the network employs a parallel detection framework and shares multi-scale layers (both high and low), enabling efficient detection of vehicles with various sizes. Then, traffic density is estimated based on the weights assigned to the categorized vehicles obtained from the improved SSD and the recorded area. Finally, traffic is classified as high, low, or moderate by comparing the estimated density to a threshold value. By adding attention characteristics of different scales to the original detection branch and replacing the VGG16 with ResNet50 of the SSD technique using our method, the feature representation capability and detection accuracy are both significantly improved.

A coarse-to-fine point completion network with details compensation and structure enhancement

Point cloud completion, the issue of estimating the complete geometry of objects from partially-scanned point cloud data, becomes a fundamental task in many 3d vision and robotics applications. To address the limitations on inadequate prediction of shape details for traditional methods, a novel coarse-to-fine point completion network (DCSE-PCN) is introduced in this work using the modules of local details compensation and shape structure enhancement for effective geometric learning. The coarse completion stage of our network consists of two branches—a shape structure recovery branch and a local details compensation branch, which can recover the overall shape of the underlying model and the shape details of incomplete point cloud through feature learning and hierarchical feature fusion. The fine completion stage of our network employs the structure enhancement module to reinforce the correlated shape structures of the coarse repaired shape (such as regular arrangement or symmetry), thus obtaining the completed geometric shape with finer-grained details. Extensive experimental results on ShapeNet dataset and ModelNet dataset validate the effectiveness of our completion network, which can recover the shape details of the underlying point cloud whilst maintaining its overall shape. Compared to the existing methods, our coarse-to-fine completion network has shown its competitive performance on both chamfer distance (CD) and earth mover distance (EMD) errors. Such as, the repairing results on the ShapeNet dataset of our completion network are reduced by an average of 35.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$35.62\\%$$\\end{document}, 33.31%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$33.31\\%$$\\end{document}, 29.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$29.62\\%$$\\end{document}, and 23.62%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$23.62\\%$$\\end{document} in terms of CD error, comparing with PCN, FoldingNet, Atlas, and CRN methods, respectively; and also reduced by an average of 15.63%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$15.63\\%$$\\end{document}, 1.29%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$1.29\\%$$\\end{document}, 64.52%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$64.52\\%$$\\end{document}, and 62.87%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$62.87\\%$$\\end{document} in terms of EMD error, respectively. Meanwhile, the completion results on the ModelNet dataset of our network have an average reduction of 28.41%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$28.41\\%$$\\end{document}, 26.57%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$26.57\\%$$\\end{document}, 20.65%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$20.65\\%$$\\end{document}, and 18.55%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$18.55\\%$$\\end{document} in terms of CD error, comparing to PCN, FoldingNet, Atlas, and CRN methods, respectively; and also an average reduction of 21.91%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$21.91\\%$$\\end{document}, 19.59%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$19.59\\%$$\\end{document}, 43.51%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$43.51\\%$$\\end{document}, and 21.49%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$21.49\\%$$\\end{document} in terms of EMD error, respectively. Our proposed point completion network is also robust to different degrees of data incompleteness and model noise.