Search

In the biomedical domain, the technologies like 3D computer vision and Bio-CAD arriving significant attention for computerized diagnosis, analysis and treatment of head and neck fractures. The advanced medical scanning devices internally scan and assemble the fragmented geometric data of the human body. The assembled skull model frequently suffers from damages caused by the process of the skull assembly process. Such damaged skull data may lead to missing some vital data for further medical analysis. Thus it is necessary to have an automatic mechanism of skull prototyping or completion before detect damaged skull models and repair them automatically for medical investigation. Automatic skull damage detection approach proposed using computer vision and machine learning methods in this paper. The input skull model in 3D format converted into 2D followed by the pre-processing operation to denoise and enhance the image quality. Then the Region of Interest (ROI) performed a dynamic binary segmentation technique. The automatic and manual features extracted from ROI using Convolutional Neural Network (CNN) layers and hybrid methods respectively. The hybrid model includes the structural, regional, and histogram features followed by its concatenation and normalization. The hybrid feature set is feed to conventional machine learning methods for skull damage detection. Automatic damage detection in input skull image is performed by the consolidated deep learning model using CNN (For features extraction) and Long-Short Term Memory (LSTM) for categorization called CNN-LSTM. The experimental outcomes show the high classification accuracy using the deep learning model compared to other machine learning techniques.

Deep learning (DL) offers an unprecedented opportunity to revolutionize the landscape of toxicity prediction based on quantitative structure-activity relationship (QSAR) studies in the big data era. However, the structural description in the reported DL-QSAR models is still restricted to the two-dimensional level. Inspired by point clouds, a type of geometric data structure, a novel three-dimensional (3D) molecular surface point cloud with electrostatic potential (SepPC) was proposed to describe chemical structures. Each surface point of a chemical is assigned its 3D coordinate and molecular electrostatic potential. A novel DL architecture SepPCNET was then introduced to directly consume unordered SepPC data for toxicity classification. The SepPCNET model was trained on 1317 chemicals tested in a battery of 18 estrogen receptor-related assays of the ToxCast program. The obtained model recognized the active and inactive chemicals at accuracies of 82.8 and 88.9%, respectively, with a total accuracy of 88.3% on the internal test set and 92.5% on the external test set, which outperformed other up-to-date machine learning models and succeeded in recognizing the difference in the activity of isomers. Additional insights into the toxicity mechanism were also gained by visualizing critical points and extracting data-driven point features of active chemicals.

The difficulty of acquiring resting-state fMRI of early developing children under the same condition leads to a dedicated protocol, i.e., scanning younger infants during sleep and older children during being awake, respectively. However, the obviously different brain activities of sleep and awake states arouse a new challenge of awake-to-sleep connectome prediction/translation, which remains unexplored despite its importance in the longitudinally-consistent delineation of brain functional development. Due to the data scarcity and huge differences between natural images and geometric data (e.g., brain connectome), existing methods tailored for image translation generally fail in predicting functional connectome from awake to sleep. To fill this critical gap, we unprecedentedly propose a novel reference-relation guided autoencoder with deep CCA restriction (R(2)AE-dCCA) for awake-to-sleep connectome prediction. Specifically, 1) A reference-autoencoder (RAE) is proposed to realize a guided generation from the source domain to the target domain. The limited paired data are thus greatly augmented by including the combinations of all the age-restricted neighboring subjects as the references, while the target-specific pattern is fully learned; 2) A relation network is then designed and embedded into RAE, which utilizes the similarity in the source domain to determine the belief-strength of the reference during prediction; 3) To ensure that the learned relation in the source domain can effectively guide the generation in the target domain, a deep CCA restriction is further employed to maintain the neighboring relation during translation; 4) New validation metrics dedicated for connectome prediction are also proposed. Experimental results showed that our proposed R(2)AE-dCCA produces better prediction accuracy and well maintains the modular structure of brain functional connectome in comparison with state-of-the-art methods.

Laser scanners can easily acquire the geometric data of physical environments in the form of point clouds. Industrial 3D reconstruction processes generally recognize objects from point clouds, which should include both geometric and semantic data. However, the recognition process is often a bottleneck in 3D reconstruction because it is labor intensive and requires expertise in domain knowledge. To address this problem, various methods have been developed to recognize objects by retrieving their corresponding models from a database via input geometric queries. In recent years, geometric data conversion to images and view-based 3D shape retrieval applications have demonstrated high accuracies. Depth images that encode the depth values as pixel intensities are frequently used for view-based 3D shape retrieval. However, geometric data collected from objects are often incomplete owing to occlusions and line-of-sight limitations. Images generated by occluded point clouds lower the view-based 3D object retrieval performance owing to loss of information. In this paper, we propose a viewpoint and image-resolution estimation method for view-based 3D shape retrieval from point cloud queries. Further, automatic selection of viewpoint and image resolution are proposed using the data acquisition rate and density calculations from sampled viewpoints and image resolutions. The retrieval performances for images generated by the proposed method are investigated via experiments and compared for various datasets. Additionally, view-based 3D shape retrieval performance with a deep convolutional neural network was investigated using the proposed method.

The 3D point clouds is an important type of geometric data structure, and the analysis of 3D point clouds based on deep learning is a very challenging task due to the disorder and irregularity. In existing research, RS-CNN provides an effective and promising method to obtain shape features on disordered point clouds directly, which encodes local features effectively. However, RS-CNN fails to consider point-wise features and global features, which are conducive to point clouds better. In this paper, we proposed PointFusionNet, which solves these problems effectively by fusing point-wise features, local features, and global features. We have designed Feature Fusion Convolution (FF-Conv) and Global Relationship Reasoning Module (GRRM) to build PointFusionNet. The point-wise features were fused with their corresponding local features in the FF-Conv and then mapped into a high-dimensional space to extract richer local features. The GRRM inferred the relationship between various parts, in order to capture global features for enriching the content of the feature descriptor. Therefore the PointFusionNet is suitable for point clouds classification and semantic segmentation by using the two distinctive modules. The PointFusionNet has been tested on ModelNet40 and ShapeNet part datasets, and the experiments show that PointFusionNet has a competitive advantage in shape classification and part segmentation tasks.

In this letter, we propose a method for accurately estimating the 6-Degree Of Freedom (DOF) pose in an urban environment when a High Definition (HD) map is available. An HD map expresses 3D geometric data with semantic information in a compressed format and thus is more memory-efficient than point cloud maps. The small capacity of HD maps can be a significant advantage for autonomous vehicles in terms of map storage and updates within a large urban area. Unfortunately, existing approaches failed to sufficiently exploit HD maps by only estimating partial pose. In this study, we present a full 6-DOF localization against an HD map using an onboard stereo camera with semantic information from roads. We introduce an 8-bit representation for road information, which allow for effective bitwise operation when matching between query data and the HD map. For the pose estimation, we leverage a particle filter followed by a full 6-DOF pose optimization. Our experimental results show a median error of approximately 0.3 m in the lateral and longitudinal directions for a drive of approximately 11 km. These results can be used by autonomous vehicles to correct the global position without using Global Positioning System (GPS) data in highly complex urban environments. The median operation speed is approximately 60 msec supporting 10 Hz.

In this paper, a parameterized automatic programming solution whose advantage depends on the automatic feature recognition of digital image is proposed and applied to the development of automatic programming software for composite grinding. This solution can overcome the difficulty of recognizing intersecting feature of complicated rotational part. And a 7-layer CNN classifier is utilized to decide whether the part has the internal features or not, which makes the proposed feature recognition method more intelligent than retrieving the unknown objects blindly. The emphasis of the research is the geometric data extraction algorithm which is the synthesis of border following algorithm, corner detection algorithm and a variety of morphological processing. Under the condition of 12000 × 12000 pixel dimension and 200-dpi resolution of input image, the relative errors between the extracted and actual values of various geometric data are all less than 0.05% for the rotational parts of maximum diameter 500 mm and maximum length of 1500 mm. And all the extracted values of geometric data rounded to integers can fully meet the requirements of NC programming. The automatic programming software based on the proposed solution has excellent portability and practicability, which is independent of any CAD tools or data exchange standards. After the programs generated by the automatic programming software are validated in simulation software NCSIMUL, the software is integrated into HNC-848 CNC system and applied in the prototype of H377 composite grinding center.

In this paper, we propose an approach to learn generic multi-modal mesh surface representations using a novel scheme for fusing texture and geometric data. Our approach defines an inverse mapping between different geometric descriptors computed on the mesh surface or its down-sampled version, and the corresponding 2D texture image of the mesh, allowing the construction of fused geometrically augmented images (FGAI). This new fused modality enables us to learn feature representations from 3D data in a highly efficient manner by simply employing standard CNNs in a transfer-learning mode. The proposed approach is both computationally and memory efficient, preserves intrinsic geometric information and learns highly discriminative feature representations by effectively fusing shape and texture information at data level. The efficacy of our approach is demonstrated for the tasks of facial action unit detection and expression classification. The extensive experiments conducted on the Bosphorus and BU-4DFE datasets show that our method produces a significant boost in the performance when compared to state-of-the-art solutions.

Multispectral LiDAR (Light Detection And Ranging) is characterized of the completeness and consistency of its spectrum and spatial geometric data, which provides a new data source for land-cover classification. In recent years, the convolutional neural network (CNN), compared with traditional machine learning methods, has made a series of breakthroughs in image classification, object detection, and image semantic segmentation due to its stronger feature learning and feature expression abilities. However, traditional CNN models suffer from some issues, such as a large number of layers, leading to higher computational cost. To address this problem, we propose a CNN-based multi-spectral LiDAR land-cover classification framework and analyze its optimal parameters to improve classification accuracy. This framework starts with the preprocessing of multi-spectral 3D LiDAR data into 2D images. Next, a CNN model is constructed with seven fundamental functional layers, and its hyper-parameters are comprehensively discussed and optimized. The constructed CNN model with the optimized hyper-parameters was tested on the Titan multi-spectral LiDAR data, which include three wavelengths of 532 nm, 1064 nm, and 1550 nm. Extensive experiments demonstrated that the constructed CNN with the optimized hyper-parameters is feasible for multi-spectral LiDAR land-cover classification tasks. Compared with the classical CNN models (i.e., AlexNet, VGG16 and ResNet50) and our previous studies, our constructed CNN model with the optimized hyper-parameters is superior in computational performance and classification accuracies.

Geometric data are commonly expressed using point clouds, with most 3D data collection devices outputting data in this form. Research on processing point cloud data for deep learning is ongoing. However, it has been difficult to apply such data as input to a convolutional neural network (CNN) or recurrent neural network (RNN) because of their unstructured and unordered features. In this study, this problem was resolved by arranging point cloud data in a canonical space through a graph CNN. The proposed graph CNN works dynamically at each layer of the network and learns the global geometric features by capturing the neighbor information of the points. In addition, by using a squeeze-and-excitation module that recalibrates the information for each layer, we achieved a good trade-off between the performance and the computation cost, and a residual-type skip connection network was designed to train the deep models efficiently. Using the proposed model, we achieved a state-of-the-art performance in terms of classification and segmentation on benchmark datasets, namely ModelNet40 and ShapeNet, while being able to train our model 2 to 2.5 times faster than other similar models.

In this paper, we propose a learning-based approach for automatic detection of hole boundary points in a 3D point cloud. Point cloud is an important geometric data structure used in 3D modelling. Data obtained from automatic acquisition techniques often result in geometric deficiencies such as holes in the 3D point cloud. For successful hole-filling to achieve better surface reconstruction, accurate detection of hole boundary points in the point cloud is necessary. Most of the existing methods use threshold values for different parameters which need to be set manually after analyzing the nature of point cloud. It becomes difficult to generalize the threshold values for a wide variety of point clouds. To overcome this limitation, we propose a deep learning framework to detect holes in the point cloud. The architecture directly consumes point cloud and considers the permutation invariance of points. Each point in the point cloud is classified as a hole boundary point or not. The detected hole boundary points are used for hole filling by fitting a surface and interpolating points on the surface.

Deep learning methods have achieved great success in analyzing traditional data such as texts, sounds, images and videos. More and more research works are carrying out to extend standard deep learning technologies to geometric data such as point cloud or voxel grid of 3D objects, real life networks such as social and citation network. Many methods have been proposed in the research area. In this work, we aim to provide a comprehensive survey of geometric deep learning and related methods. First, we introduce the relevant knowledge and history of geometric deep learning field as well as the theoretical background. In the method part, we review different graph network models for graphs and manifold data. Besides, practical applications of these methods, datasets currently available in different research area and the problems and challenges are also summarized.

Machine Learning (ML) techniques are now being used very successfully in predicting and supporting decisions in multiple areas such as environmental issues and land management. These techniques have also provided promising results in the field of natural hazard assessment and risk mapping. The aim of this work is to apply the Supervised ML technique to train a model able to classify a particular gravity-driven coastal hillslope geomorphic model (slope-over-wall) involving most of the soft rocks of Cilento (southern Italy). To train the model, only geometric data have been used, namely morphometric feature maps computed on a Digital Terrain Model (DTM) derived from Light Detection and Ranging (LiDAR) data. Morphometric maps were computed using third-order polynomials, so as to obtain products that best describe landforms. Not all morphometric parameters from literature were used to train the model, the most significant ones were chosen by applying the Neighborhood Component Analysis (NCA) method. Different models were trained and the main indicators derived from the confusion matrices were compared. The best results were obtained using the Weighted k-NN model (accuracy score = 75%). Analysis of the Receiver Operating Characteristic (ROC) curves also shows that the discriminating capacity of the test reached percentages higher than 95%. The model, resulting more accurate in the training area, will be extended to similar areas along the Tyrrhenian coastal land.

Deep learning for 3D data has become a popular research theme in many fields. However, most of the research on 3D data is based on voxels, 2D images, and point clouds. At actual industrial sites, face-based geometry data are being used, but their direct application to industrial sites remains limited due to a lack of existing research. In this study, to overcome these limitations, we present a face-based variational autoencoder (FVAE) model that generates 3D geometry data using a variational autoencoder (VAE) model directly from face-based geometric data. Our model improves the existing node and edge-based adjacency matrix and optimizes it for geometric learning by using a face- and edge-based adjacency matrix according to the 3D geometry structure. In the experiment, we achieved the result of generating adjacency matrix information with 72% precision and 69% recall through end-to-end learning of Face-Based 3D Geometry. In addition, we presented various structurization methods for 3D unstructured geometry and compared their performance, and proved the method to effectively perform reconstruction of the learned structured data through experiments.

Today one of the most important problems in modern Augmented and Mixed reality applications is the analysis of surrounding objects in order to trace variability of the observed environment, the behavior of which is usually unpredictable. The solution to this problem can be widely applied in many mobile applications as a tool to interact with surrounding objects in the environment. Augmented and Mixed realities are both extremely promising directions for further research of interaction with the real environment. In this paper, we suggest a method of creation of geometric data based on a series of points in order to reconstruct the surface of real objects. This allows us to ensure the interaction with virtual objects. The proposed method comprises three steps: point detection, points clusterization into multiple groups of points, depending on their location, and the creation of geometric data such as lines and surfaces.

At present, inspection systems process visual data captured by cameras, with deep learning approaches applied to detect defects. Defect detection results usually have an accuracy higher than 94%. Real-life applications, however, are not very common. In this paper, we describe the development of a tire inspection system for the tire industry. We provide methods for processing tire sidewall data obtained from a camera and a laser sensor. The captured data comprise visual and geometric data characterizing the tire surface, providing a real representation of the captured tire sidewall. We use an unfolding process, that is, a polar transform, to further process the camera-obtained data. The principles and automation of the designed polar transform, based on polynomial regression (i.e., supervised learning), are presented. Based on the data from the laser sensor, the detection of abnormalities is performed using an unsupervised clustering method, followed by the classification of defects using the VGG-16 neural network. The inspection system aims to detect trained and untrained abnormalities, namely defects, as opposed to using only supervised learning methods.

This work presents a user-study evaluation of various visual and haptic feedback modes on a real telemanipulation platform. Of particular interest is the potential for haptic guidance virtual fixtures and 3D-mapping techniques to enhance efficiency and awareness in a simple teleoperated valve turn task. An RGB-Depth camera is used to gather real-time color and geometric data of the remote scene, and the operator is presented with either a monocular color video stream, a 3D-mapping voxel representation of the remote scene, or the ability to place a haptic guidance virtual fixture to help complete the telemanipulation task. The efficacy of the feedback modes is then explored experimentally through a user study, and the different modes are compared on the basis of objective and subjective metrics. Despite the simplistic task and numerous evaluation metrics, results show that the haptic virtual fixture resulted in significantly better collision avoidance compared to 3D visualization alone. Anticipated performance enhancements were also observed moving from 2D to 3D visualization. Remaining comparisons lead to exploratory inferences that inform future direction for focused and statistically significant studies.

This paper inaugurates a brand new five-axis tool path generation method for freeform surface machining, which is solely based on digital image processing. By utilizing image representation, a freeform surface can be digitized into a uniform grayscale image. With this novel representation, we come up with a region based five-axis tool path scheme with remarkably enhanced utility towards freeform surface machining process. The image representation possesses sufficient geometric data for us to evaluate the surface normal distribution and split the surface area accordingly via image processing algorithms. Since each partitioned region contains a limited surface inclination range, one rotary axis is locked with a certain angle during the processing of this region. In addition, by further exploiting the potential of the image representation, the cutter contact (CC) curves are extracted purely from the image of each partitioned region. Preliminary results show that this method is more than adequate for the tool path computation. The computed tool path outperforms the benchmarks by reducing the machining time for as much as 40%, as verified in the physical cutting test.

CORTEX is a cognitive robotics architecture inspired by three key ideas: modularity, internal modelling and graph representations. CORTEX is also a computational framework designed to support early forms of intelligence in real world, human interacting robots, by selecting an a priori functional decomposition of the capabilities of the robot. This set of abilities was then translated to computational modules or agents, each one built as a network of software interconnected components. The nature of these agents can range from pure reactive modules connected to sensors and/or actuators, to pure deliberative ones, but they can only communicate with each other through a graph structure called Deep State Representation (DSR). DSR is a short-term dynamic representation of the space surrounding the robot, the objects and the humans in it, and the robot itself. All these entities are perceived and transformed into different levels of abstraction, ranging from geometric data to high-level symbolic relations such as "the person is talking and gazing at me". The combination of symbolic and geometric information endows the architecture with the potential to simulate and anticipate the outcome of the actions executed by the robot. In this paper we present recent advances in the CORTEX architecture and several real-world human-robot interaction scenarios in which they have been tested. We describe our interpretation of the ideas inspiring the architecture and the reasons why this specific computational framework is a promising architecture for the social robots of tomorrow.

The increasing accessibility of low-cost, desktop scanners has provided researchers with the means to acquire 3D geometric data that can be used to create digital archives of archaeological artifacts, such as bones, for geometric morphometric study in the fields of osteology and forensic anthropology. However, the process of digitizing and assessing in silico models is not straight forward. In this two-part study, we first compared the performance of the 2010 NextEngine model 2020i laser scanner (NE) and Hewlett-Packard Pro S3 structured light scanner (HP) to that of the Keyence VHX 6000 digital microscope (VHX), using a fabricated test block containing bas-relief features with offsets ranging from 3 mm down to 0.1 mm. The xyz coordinate data from the in silico models were also compared to xyz coordinate data obtained using a Microscribe MX 3D coordinate measurement machine (CMM), which showed no evidence of geometric distortion in the in silico models. We applied Bland-Altman analysis to quantify the coefficient of repeatability (CoR) for each instrument used in the study, and to determine agreement between the scanners and the VHX. Dimensional accuracy was assessed by comparing bas-relief offset dimensions on the in silico models derived from the scanners to the VHX measurements of the same features. Second, as an application to osteology we examined intra-observer scanning protocol variability, using the coefficient of variation (CV) to quantify surface area and volume variance between repeated scans of eight porcine capital femoral epiphyses with undulating mammillary processes on one surface with amplitudes covering the range of the test block bas-relief offset values. The CoR showed each test-retest measurement from each instrument differed by no more than their CoR: 0.010 mm, 0.137 mm, 0.068 mm, 0.193 mm for the VHX, NE, HP and CMM, respectively. There was agreement between the instruments, but each instrument (NE, HP and CMM) overestimated bas-relief features as reported by the VHX, on average (bias) by 0.046 ± 0.038 mm, 0.025 ± 0.033 mm, 0.026 ± 0.033 mm for the NE, HP and CMM, respectively. Both scanners captured surface features as small as 0.1 mm. For the intra-observer scanning protocol of the epiphyseal bones using the NE, the coefficient of variation (mean CV ± SD) was 0.69 ± 0.25% for the bone surface area and 0.77 ± 0.26% for bone volume. The scanner appears suitable for capturing bas-relief features in the range of 3 to 0.1 mm, while preserving the ratios of distances between measured points within the spatial volume surveyed. Examples of potential applications include scanning bone or pottery fragments for virtual reconstruction, where capturing dimensional detail and areas is important, reproducing the fine surface features of pottery, figurines, etc. for cultural preservation and for examination by scientists who cannot study the objects directly, as well as to provide access to the public by electronic means. The lower cost of these scanners may provide museums with small budgets with a means to begin to archive some items in their collection.

Sweep scan is an emerging five-axis inspection technology of retrieving geometric data from free-form surfaces. Compared with the traditional surface inspection by coordinate measuring machine (CMM) that works in a point-by-point manner, sweep scan is able to make the stylus tip continuously sweep at high speed on the surface to inspect, whilst acquiring the 3-D point data accurately and efficiently. However, at present, for an arbitrary free-form surface, it still heavily depends on humans to manually generate a suitable sweep scan path. This paper presents a practical sweep scan path planning method, which can automatically generate an efficient sweep scan path for an arbitrary free-form surface. The resultant sweep scan path is able to cater to the shape of the surface to inspect and the unique kinematic characteristics of the typical five-axis inspection machine, so as to achieve a tremendous increase in inspection efficiency. The experiments performed by us show that when compared with some existing automatic sweep scan path generation algorithms such as the simple zigzag method, the proposed method is able to reduce the total inspection time by as much as seven times. Note to Practitioners —This paper is motivated by the lack of effective automatic sweep scan path planning methods for arbitrary free-form surfaces. Although there is a rich body of five-axis CMM path planning methods for inspection of free-form surfaces, they are mainly designed for the traditional inspection mode, i.e., the point-by-point measurement. Compared with the traditional discrete way of CMM inspection, sweep scan has a quite different inspection mode, which tremendously boosts the inspection efficiency by maintaining the constant contact between the stylus and the surface to inspect. The traditional five-axis inspection path planning methods are no longer applicable to sweep scan. Up to date, only a very handful few of automatic sweep scan path planning methods are reported and they seem to be very difficult to implement and no optimality of inspection efficiency is discussed. We present a practical algorithm for automatic generation of sweep scan paths for inspection of an arbitrary free-form surface. The algorithm is novel in that, by considering both the surface shape and the unique kinematic characteristics of a typical five-axis inspection machine, it tries to globally minimize the total inspection time. In our experiments, the sweep scan paths, generated by the proposed method, are found to outperform that of some traditional simple sweep scan path generation methods such as the zigzag method by a large margin in terms of total inspection time (while the data acquisition accuracy is maintained at the same level). The method is particularly suitable for those industrial parts that are large but relatively smooth, such as windmill blades and automobile panels.

Made available in DSpace on 2019-10-06T15:26:36Z (GMT). No. of bitstreams: 0 Previous issue date: 2019-01-01 Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) Academy of Finland Miniaturised 2D frame format hyperspectral camera technology that is suitable for small unmanned aerial vehicles (UAVs) has entered the market, making the generation of hyperspectral digital surface models (HDSMs) feasible. HDSMs offer a rigorous approach to capturing the target spectral and 3D geometric data. The main objective of this investigation was to study and develop techniques for the generation of HDSMs in forest areas using novel hyperspectral 2D frame camera technologies. An approach based on object-space image matching was developed, adapting the traditional vertical line locus (VLL) method for HDSM generation; this was then named the hyperspectral VLL (HVLL) approach. Additionally, image classification was introduced into the processing chain in order to adapt the matching parameters, based on different classes. We also proposed a method for extracting the spectral and viewing angle information of the points. An empirical study was carried out using UAV datasets from tropical and boreal forests using 2D format hyperspectral cameras, based on tuneable Fabry-Pérot interferometer (FPI) technology. Quality assessment was performed using DSMs based on state-of-the-art commercial software and airborne laser scanning (ALS). The results showed that the proposed technique generated a high-quality HDSM in both tested environments. The HDSM had higher deviations over the continuous canopy cover than the digital surface models (DSMs) generated using commercial software. The method using image classification information outperformed the commercial approach with respect to the ability to measure ground points in shadowed areas and in canopy gaps. The proposed method is of great interest in supporting automated interpretations of novel multi- and hyperspectral imaging technologies, especially when applied complex objects, such as forests. UNESP São Paulo State University Finnish Geospatial Research Institute, Masala, Kirkkonummi UNESP São Paulo State University FAPESP: 2013/17787-3 FAPESP: 2013/50426-4 FAPESP: 2014/24844-6 Academy of Finland: 273806

Accurate capture of digital data from a manufactured part is an essential source of information about the quality of the produced part and the health of the corresponding manufacturing processes. The commonly known precise instruments for digitizing work-pieces on Coordinate Metrology Machine (CMM) are contact probes. Although, contact techniques confront difficulties, they lead to higher precisions comparing to non-contact methods. Ball-tip-radius error is one of the significant faults of contact methods. Conventional error compensating approaches of ball-tip-radius error often demand complicated and time consuming post-processes. Furthermore, considering the limitations of typical serial structure in commercial CMMs, adaptive accurate path planning is a very complex task. This paper presents an intelligent combination of a novel 4-DOF parallel mechanism implementing, a linear cost-effective contact probe, and an adaptive path planning method to automatically correct the probing direction. The implemented parallel structure has three translational DOF in X, Y and Z directions and one rotational DOF around X axis. A precise turn-table provides rotation of work-piece around the Z axis. The developed algorithm utilizes a data clustering approach and intelligently generates an automated adaptive probe placement and motion path. Validation process using simulation and actual measurement of free-form surfaces demonstrate highly reliable feasibility for eliminating the ball-tip-radius error resulting in high precision 3D data capturing process.

This article proposes an active learning method for high dimensional data, based on intrinsic data geometries learned through diffusion processes on graphs. Diffusion distances are used to parametrize low-dimensional structures on the dataset, which allow for high-accuracy labelings of the dataset with only a small number of carefully chosen labels. The geometric structure of the data suggests regions that have homogeneous labels, as well as regions with high label complexity that should be queried for labels. The proposed method enjoys theoretical performance guarantees on a general geometric data model, in which clusters corresponding to semantically meaningful classes are permitted to have nonlinear geometries, high ambient dimensionality, and suffer from significant noise and outlier corruption. The proposed algorithm is implemented in a manner that is quasilinear in the number of unlabeled data points, and exhibits competitive empirical performance on synthetic datasets and real hyperspectral remote sensing images., Comment: 20 pages, 10 figures

Dense image matching using remote sensing images is of particular importance for providing ground geometric data for object modeling and analysis. These methods normally operate in a rectified (epipolar) space producing disparity images that are correlated with the final three-dimensional geometry. However, the resulting digital surface models can be problematic on discontinuities of the terrain object, e.g., building edges, which largely limit their practical use for highly accurate object modeling. Existing works put forth efforts for dealing with this issue by defining better edge constraints and more global energy optimization, while intended to improve the disparity maps, it might be challenging to leverage the result of all the pixels through a single energy minimization, leading to either oversmoothed object boundaries or noisy surfaces. In this paper, we propose an intuitive method that integrates straight line primitives to enhance the disparity map. For each matched line, we perform a local discontinuity analysis and propose an intensity-based weighting method for a local plane fitting using iteratively solved weighted least squares adjustment, such that straightness of the object's edges (e.g., buildings) can be preserved, as the straight lines are detected and matched. Experiments on both aerial and satellite dataset show that the proposed method yields visually clear object edges and has achieved higher accuracy (2–3 pixels improvement) around edges than DSM generated from typical stereo matching result (i.e., from semiglobal matching).

Graphs frequently appear as a type of data structure that efficiently models a set of interrelated objects as nodes and their relations as edges between them. Geometric deep learning emerged as a promising field for modeling non-Euclidean geometric data as graphs, Riemannian manifolds and meshes. In this paper, we review geometric deep learning based state-of-the-art approaches for multi-object tracking. We illustrate the main concept, the model architecture, loss functions, and optimization strategies of the main algorithms. Moreover, algorithms are quantitatively evaluated on standard performance metrics using the standard dataset. The potential of geometric deep learning in tracking is discussed and future directions are proposed.

The drone navigation requires the comprehensive understanding of both visual and geometric information in the 3D world. In this paper, we present a Visual-Geometric Fusion Network(VGF-Net), a deep network for the fusion analysis of visual/geometric data and the construction of 2.5D height maps for simultaneous drone navigation in novel environments. Given an initial rough height map and a sequence of RGB images, our VGF-Net extracts the visual information of the scene, along with a sparse set of 3D keypoints that capture the geometric relationship between objects in the scene. Driven by the data, VGF-Net adaptively fuses visual and geometric information, forming a unified Visual-Geometric Representation. This representation is fed to a new Directional Attention Model(DAM), which helps enhance the visual-geometric object relationship and propagates the informative data to dynamically refine the height map and the corresponding keypoints. An entire end-to-end information fusion and mapping system is formed, demonstrating remarkable robustness and high accuracy on the autonomous drone navigation across complex indoor and large-scale outdoor scenes. The dataset can be found in http://vcc.szu.edu.cn/research/2021/VGFNet., Accepted by CVM 2021

Machine learning models are known to be vulnerable to adversarial attacks, namely perturbations of the data that lead to wrong predictions despite being imperceptible. However, the existence of "universal" attacks (i.e., unique perturbations that transfer across different data points) has only been demonstrated for images to date. Part of the reason lies in the lack of a common domain, for geometric data such as graphs, meshes, and point clouds, where a universal perturbation can be defined. In this paper, we offer a change in perspective and demonstrate the existence of universal attacks for geometric data (shapes). We introduce a computational procedure that operates entirely in the spectral domain, where the attacks take the form of small perturbations to short eigenvalue sequences; the resulting geometry is then synthesized via shape-from-spectrum recovery. Our attacks are universal, in that they transfer across different shapes, different representations (meshes and point clouds), and generalize to previously unseen data., Published at CVPR 2021

Turnouts are the most expensive part to maintain on the railway track and therefore automated systems for detecting turnout defects are of great interest. Machine learning can improve predictive maintenance and is often used in automatic systems for precise prognosis. In this study, machine learning is used for identifying the condition of railway turnouts and potentially reducing costs by early automatic detection of defects. To train a machine learning algorithm, ordered, structured and categorized data (labelled data) are needed. A method is proposed to label the condition of turnouts in the Danish Railway based on a collection of repair records. This labelling of the turnouts is accomplished with unsupervised methods, namely a principal component analysis (PCA) followed by a cluster analysis. The labelling of the turnouts is investigated through comparisons of geometric measurements captured from the recording car. The difference in the physical properties illustrated by the geometric data indicates that the labelling is a good indicator of the relative condition of the turnout. When the data are labelled, supervised learning can be used to optimize the predictive power of machine learning algorithms (i.e. the algorithm learns from the labelled data) for classification of turnouts.

Emotional recognition from videos or images requires large amount of data to obtain high performance and classification accuracy. However, large datasets are not always easily available. A good solution to this problem is to augment the data and extrapolate it to create a bigger dataset for training the classifier. In this paper, we evaluate the impact of different geometric data augmentation (GDA) techniques on emotion recognition accuracy using facial image data. The GDA techniques that were implemented were horizontal reflection, cropping, rotation separately and combined. In addition to this, our system was further evaluated with four different classifiers (Convolutional Neural Network (CNN), Linear Discriminant Analysis (LDA), K-Nearest Neighbor (kNN) and Decision Tree (DT)) to determine which of the four classifiers achieves the best results. In the proposed system, we used augmented data from a dataset (SAVEE) to perform training, and testing was carried out by the original data. A combination of GDA techniques using the CNN classifier was found to give the best performance of approximately 97.8%. Our system with GDA augmentation was shown to outperform previous approaches where only the original dataset was used for classifier training.

Deep convolutional networks have become the mainstream in computer vision applications. Although CNNs have been successful in many computer vision tasks, it is not free from drawbacks. The performance of CNN is dramatically degraded by geometric transformation, such as large rotations. In this paper, we propose a novel CNN architecture that can improve the robustness against geometric transformations without modifying the existing backbones of their CNNs. The key is to enclose the existing backbone with a geometric transformation (and the corresponding reverse transformation) and a feature map ensemble. The proposed method can inherit the strengths of existing CNNs that have been presented so far. Furthermore, the proposed method can be employed in combination with state-of-the-art data augmentation algorithms to improve their performance. We demonstrate the effectiveness of the proposed method using standard datasets such as CIFAR, CUB-200, and Mnist-rot-12k., Comment: Accepted to ICIP2021

During each cycle of automotive development, large amounts of geometric data are generated as results of design studies and simulation tasks. Discovering hidden knowledge from this data and making it available to the development team strengthens the design process by utilizing historic information when creating novel products. To this end, we propose to use powerful geometric deep learning models that learn lowdimensional representation of the design data in an unsupervised fashion. Trained models allow to efficiently explore the design space, as well as to generate novel designs. One popular class of generative models are variational autoencoders, which have however been rarely applied to geometric data. Hence, we use a variational autoencoder for 3D point clouds (PC-VAE) and explore the model’s generative capabilities with a focus on the generation of realistic yet novel 3D shapes. We apply the PC-VAE to point clouds sampled from car shapes from a benchmark data set and employ quantitative measures to show that our PC-VAE generates realistic car shapes, wile returning a richer variety of unseen shapes compared to a baseline autoencoder. Finally, we demonstrate how the PC-VAE can be guided towards generating shapes with desired target properties by optimizing the parameters that maximize the output of a trained classifier for said target properties. We conclude that generative models are a powerful tool that may aid designers in automotive product development.

Among all possible postures of the human body, NASA selected eight of principle importance for space exploration. The current article aims to determine the mass-inertial characteristics of the human body of the average Bulgarian male in one of these positions—the so-called relaxed, or weightless position. We determine the corresponding characteristics of the centre of mass and principal moments of inertia using a 16-segmental biomechanical mathematical model that is generated within the SolidWorks environment. We verify the model by comparing the analytical results for each of the body segments with the results obtained using the computer model. The geometric data needed for the construction of the 3D model are taken to be in correspondence with experimentally available anthropometric data for about 2500 Bulgarian men. One their basis one determines the characteristics of the average Bulgarian men. Then, using the CAD realization of the model the inertial parameters of this “average” male in different body positions can be determined. The comparison made between our model results described in this article and the data reported in the literature, where available, gives us confidence that the suggested model can be used to calculate the properties in question at any postures of the body of interest. In principle, our approach can be also used to calculate the corresponding mass inertial data for any individual provided that the anthropometric set of parameters for that individual are measured. The model we used is suitable when one needs such parameters in problems appearing not only in space exploration with the participation of male astronauts but also in rehabilitation, sport, criminology, robotics, etc.

During the last few years, in the field of computer vision, sophisticated deep learning methods have been developed to accomplish semantic segmentation tasks of 3D point cloud data. Additionally, many researchers have extended the applicability of these methods, such as PointNet or PointNet++, beyond semantic segmentation tasks of indoor scene data to large-scale outdoor scene data observed using airborne laser scanning systems equipped with light detection and ranging (LiDAR) technology. Most extant studies have only investigated geometric information (x, y, and z or longitude, latitude, and height) and have omitted rich radiometric information. Therefore, we aim to extend the applicability of deep learning-based model from the geometric data into radiometric data acquired with airborne full-waveform LiDAR without converting the waveform into 2D images or 3D voxels. We simultaneously train two models: a local module for local feature extraction and a global module for acquiring wide receptive fields for the waveform. Furthermore, our proposed model is based on waveform-aware convolutional techniques. We evaluate the effectiveness of the proposed method using benchmark large-scale outdoor scene data. By integrating the two outputs from the local module and the global module, our proposed model had achieved higher mean recall value 0.92 than previous methods and higher F1 scores for all six classes than the other 3D Deep Learning method. Therefore, our proposed network consisting of the local and global module successfully resolves the semantic segmentation task of full-waveform LiDAR data without requiring expert knowledge.

Visual localization is a fundamental problem in computer vision and robotics. Recently, deep learning has shown to be effective for robust monocular localization. Most deep learning-based methods utilize convolution neural network (CNN) to regress global 6 degree-of-freedom (Dof) pose. However, these methods suffer from pose sparsity, leading to over-fitting during training and poor localization performance on unseen data. In this paper, we try to alleviate this issue by implementing randomly geometric augmentation (RGA) during training. Specifically, we firstly estimate the depth map using a depth estimation network for the initial training image. Combing the estimated depth, RGB image and its corresponding pose, we can randomly synthesize new images of different views. The synthesized and initial images are used to train the pose regression network. Experiment results show our geometric augmentation strategy can significantly improve the localization accuracy.

As one of the most fundamental geometric data types for the representation of space and object shapes, a point cloud usually maintains much structural information about the spatial relationship between objects and their features. However, the relative sparseness of point clouds sampled in most practical applications make extracting information-rich features a major challenge. Traditionally, feature extraction algorithms resorted to structured feature engineering and used handcrafted representations for some specific problems. Motivated by the development of deep neural networks, many researchers started to handle the unstructured point clouds from the raw data samples of 3D scanning devices. Some important advantages that deep learning frameworks have over traditional feature engineering is generalizing complex features and associated semantic concepts in a hierarchical manner. Deep learning models have achieved significant landmarks in cognitive processing of speech, image, and video signals. However, unlike in 2D image processing, a 3D point cloud is irregular and sparse. Hence, traditional network frameworks are difficult to apply on 3D geometric data directly. In this paper, we propose to integrate a local point convolution network with spectral pooling to aggregate and learn features in 3D point clouds. The benefits of our framework are fast convergence and competitive performance on point cloud classification.

Vascular disease is a leading cause of death world wide and therefore the treatment thereof is critical. Understanding and classifying the types and levels of stenosis can lead to more accurate and better treatment of vascular disease. Some clinical techniques to measure stenosis from real patient data are invasive or of low accuracy.In this paper, we propose a new methodology, which can serve as a supplementary way of diagnosis to existing methods, to measure the degree of vascular disease using topological data analysis. We first proposed the critical failure value, which is an application of the 1-dimensional homology group to stenotic vessels as a generalization of the percent stenosis. We demonstrated that one can take important geometric data including size information from the persistent homology of a topological space. We conjecture that we may use persistent homology as a general tool to measure stenosis levels for many different types of stenotic vessels.We also proposed the spherical projection method, which is meant to allow for future classification of different types and levels of stenosis. We showed empirically using the spectral approximation of different vasculatures that this projection could provide a new medical index that measures the degree of vascular disease. Such a new index is obtained by calculating the persistence of the 2-dimensional homology of flows. We showed that the spherical projection method can differentiate between different cases of flows and reveal hidden patterns about the underlying blood flow characteristics, that is not apparent in the raw data. We showed that persistent homology can be used in conjunction with this technique to classify levels of stenosis.The main interest of this paper is to focus on the theoretical development of the framework for the proposed method using a simple set of vascular data.

Point cloud is an important type of geometric data structure. The uneven distribution of points brings challenges to the research of deep learning on point clouds. Many researchers are aware of this problem but have yet to come up with a specific solution. In this paper, we proposed a novel approach for learning receptive fields adapted to local point density variation and a density-targeted data augmentation strategy for point clouds. The receptive fields of the network can be adaptively adjusted by learning a weight matrix of local neighborhood points. Thus, our network is robust with respect to uneven distribution of points. Experiments show that methord in this paper achieves on par or better performance than state-of-the-art methods on multiple challenging benchmark datasets.

Background There is a growing need for analyzing medical data such as brain connectomes. However, the unavailability of large-scale training samples increases risks of model over-fitting. Recently, deep learning (DL) architectures quickly gained momentum in synthesizing medical data. However, such frameworks are primarily designed for Euclidean data (e.g., images), overlooking geometric data (e.g., brain connectomes). A few existing geometric DL works that aimed to predict a target brain connectome from a source one primarily focused on domain alignment and were agnostic to preserving the connectome topology. New method To address the above limitations, firstly, we adapt the graph translation generative adversarial network (GT GAN) architecture to brain connectomic data. Secondly, we extend the baseline GT GAN to a cyclic graph translation (CGT) GAN, allowing bidirectional brain network translation between the source and target views. Finally, to preserve the topological strength of brain regions of interest (ROIs), we impose a topological strength constraint on the CGT GAN learning, thereby introducing CGTS GAN architecture. Comparison with existing methods We compared CGTS with graph translation methods and its ablated versions. Results Our deep graph network outperformed the baseline comparison method and its ablated versions in mean squared error (MSE) using multiview autism spectrum disorder connectomic dataset. Conclusion We designed a topology-aware bidirectional brain connectome synthesis framework rooted in geometric deep learning, which can be used for data augmentation in clinical diagnosis.

We propose DualConvMesh-Nets (DCM-Net) a family of deep hierarchical convolutional networks over 3D geometric data that combines two types of convolutions. The first type, geodesic convolutions, defines the kernel weights over mesh surfaces or graphs. That is, the convolutional kernel weights are mapped to the local surface of a given mesh. The second type, Euclidean convolutions, is independent of any underlying mesh structure. The convolutional kernel is applied on a neighborhood obtained from a local affinity representation based on the Euclidean distance between 3D points. Intuitively, geodesic convolutions can easily separate objects that are spatially close but have disconnected surfaces, while Euclidean convolutions can represent interactions between nearby objects better, as they are oblivious to object surfaces. To realize a multi-resolution architecture, we borrow well-established mesh simplification methods from the geometry processing domain and adapt them to define mesh-preserving pooling and unpooling operations. We experimentally show that combining both types of convolutions in our architecture leads to significant performance gains for 3D semantic segmentation, and we report competitive results on three scene segmentation benchmarks. Our models and code are publicly available., Comment: CVPR 2020 camera-ready version

Establishing correspondence between shapes is a very important and active research topic in many domains. Due to the powerful ability of deep learning on geometric data, lots of attractive results have been achieved by convolutional neural networks (CNNs). In this paper, we propose a novel architecture for shape correspondence, termed Anisotropic Chebyshev spectral CNNs (ACSCNNs), based on a new extension of the manifold convolution operator. The extended convolution operators aggregate the local features of signals by a set of oriented kernels around each point, which allows to much more comprehensively capture the intrinsic signal information. Rather than using fixed oriented kernels in the spatial domain in previous CNNs, in our framework, the kernels are learned by spectral filtering, based on the eigen-decompositions of multiple Anisotropic Laplace-Beltrami Operators. To reduce the computational complexity, we employ an explicit expansion of the Chebyshev polynomial basis to represent the spectral filters whose expansion coefficients are trainable. Through the benchmark experiments of shape correspondence, our architecture is demonstrated to be efficient and be able to provide better than the state-of-the-art results in several datasets even if using constant functions as inputs.

Some face recognition methods are designed to utilize geometric information extracted from depth sensors to overcome the weaknesses of single-image based recognition technologies. However, the accurate acquisition of the depth profile is an expensive and challenging process. Here, we introduce a novel method that learns to recognize faces from stereo camera systems without the need to explicitly compute the facial surface or depth map. The raw face stereo images along with the location in the image from which the face is extracted allow the proposed CNN to improve the recognition task while avoiding the need to explicitly handle the geometric structure of the face. This way, we keep the simplicity and cost efficiency of identity authentication from a single image, while enjoying the benefits of geometric data without explicitly reconstructing it. We demonstrate that the suggested method outperforms both existing single-image and explicit depth based methods on large-scale benchmarks, and even capable of recognize spoofing attacks. We also provide an ablation study that shows that the suggested method uses the face locations in the left and right images to encode informative features that improve the overall performance.

This paper summarises a Doctoral Thesis which proposes a new approach for large- scale forest visualisation with geometrically diverse trees. The main contribution of the proposed method is an interactive visualisation of numerous trees without generating geometric data in advance, which is achieved by a new method for on-the-fly tree skeleton synthesis with a specific level of detail, and by a new procedural volumetric tree crown visualisation which avoids geometry formation altogether. The proposed method enables visualisation of forests with millions of trees, thus allowing rendering more trees than geometry-based visualisation methods.

Developing predictive intelligence in neuroscience for learning how to generate multimodal medical data from a single modality can improve neurological disorder diagnosis with minimal data acquisition resources. Existing deep learning frameworks are mainly tailored for images, which might fail in handling geometric data (e.g., brain graphs). Specifically, predicting a target brain graph from a single source brain graph remains largely unexplored. Solving such problem is generally challenged with domain fracturecaused by the difference in distribution between source and target domains. Besides, solving the prediction and domain fracture independently might not be optimal for both tasks. To address these challenges, we unprecedentedly propose a Learning-guided Graph Dual Adversarial Domain Alignment (LG-DADA) framework for predicting a target brain graph from a source brain graph. The proposed LG-DADA is grounded in three fundamental contributions: (1) a source data pre-clustering step using manifold learning to firstly handle source data heterogeneity and secondly circumvent mode collapse in generative adversarial learning, (2) a domain alignment of source domain to the target domain by adversarially learning their latent representations, and (3) a dual adversarial regularization that jointly learns a source embedding of training and testing brain graphs using two discriminators and predict the training target graphs. Results on morphological brain graphs synthesis showed that our method produces better prediction accuracy and visual quality as compared to other graph synthesis methods.

The human body is one of the most complicated objects to model because of its complex features, non-rigidity, and the time required to take body measurements. Basic technologies available in this field range from small and low-cost scanners that must be moved around the body to large and high-cost scanners that can capture all sides of the body simultaneously. This paper presents an image-based scanning system which employs the structure-from-motion method. The design and development process of the scanner includes its physical structure, electronic components, and the algorithms used for extracting 3D data. In addition to the accuracy, which is one of the main parameters to consider when choosing a 3D scanner, the time and cost of the system are among the most important parameters for evaluating a scanner system in the field of human scanning. Because of the non-static nature of the human body, the scanning time is particularly important. On the other hand, a high-cost system may lead to limited use of such systems. The design developed in this paper, which utilizes 100 cameras, facilitates the acquisition of geometric data in a fraction of a second (0.001 s) and provides the capabilities of large, freestanding scanners at a price akin to that of smaller, mobile ones., Journal of Imaging, 6 (4), ISSN:2313-433X

In this paper, we propose a novel multi-modal mesh surface representation fusing texture and geometric data. Our approach defines an inverse mapping between different geometric descriptors computed on the mesh surface or its down-sampled version, and the corresponding 2D texture image of the mesh, allowing the construction of fused geometrically augmented images. This new fused modality enables us to learn feature representations from 3D data in a highly efficient manner by simply employing standard convolutional neural networks in a transfer-learning mode. In contrast to existing methods, the proposed approach is both computationally and memory efficient, preserves intrinsic geometric information and learns highly discriminative feature representation by effectively fusing shape and texture information at data level. The efficacy of our approach is demonstrated for the tasks of facial action unit detection, expression classification, and skin lesion classification, showing competitive performance with state of the art methods.

Online semantic 3D segmentation in company with real-time RGB-D reconstruction poses special challenges such as how to perform 3D convolution directly over the progressively fused 3D geometric data, and how to smartly fuse information from frame to frame. We propose a novel fusion-aware 3D point convolution which operates directly on the geometric surface being reconstructed and exploits effectively the inter-frame correlation for high quality 3D feature learning. This is enabled by a dedicated dynamic data structure which organizes the online acquired point cloud with global-local trees. Globally, we compile the online reconstructed 3D points into an incrementally growing coordinate interval tree, enabling fast point insertion and neighborhood query. Locally, we maintain the neighborhood information for each point using an octree whose construction benefits from the fast query of the global tree.Both levels of trees update dynamically and help the 3D convolution effectively exploits the temporal coherence for effective information fusion across RGB-D frames., Comment: CVPR 2020

International audience; In this document we describe our method and the results obtained for comparing jaws of prehistoric humans. Our main goal was twofold: 1) establish a methodology for comparing the structure of 3D shapes of scans of jaws using geometric data analysis techniques, and 2) use this methodology for comparing and clustering individual objects according to their geometric similarity. Moreover, we also applied geometric modeling techniques to establish a "clean" version of the dataset, without significant artefacts present in the original data, such as large missing parts. We then applied our analysis techniques both on the original and the clean dataset in order to validate our comparison results. For our core task of shape comparison, we used recent state-of-the-art shape matching methods and we present similarity results using different comparison metrics.; Dans ce document, nous décrivons notre méthode et les résultats des comparaisons de mâchoires des humains préhistoriques. Notre objectif était double : (1) établir une méthodologie pour comparer la structure des formes 3D de scans de mâchoires à l’aide de techniques d’analyse de données géométriques ; et (2) utiliser cette méthodologie pour comparer et regrouper des individus en fonction de leur similitude géométrique. Nous avons également appliqué des techniques de modélisation géométrique pour établir une version « propre » de l’ensemble de données, sans artefact significatif dans les données d’origine. Nous avons ensuite appliqué nos techniques d’analyse à la fois sur le jeu de données d’origine et sur le jeu de données « propre » afin de valider nos résultats de comparaison. Pour la comparaison des formes, nous avons utilisé des méthodes de correspondance de formes classiques et nous présentons les similarités en utilisant différentes mesures de comparaison.

With the emergence of new intelligent sensing technologies such as 3D scanners and stereo vision, high-quality point clouds have become very convenient and lower cost. The research of 3D object recognition based on point clouds has also received widespread attention. Point clouds are an important type of geometric data structure. Because of its irregular format, many researchers convert this data into regular three-dimensional voxel grids or image collections. However, this can lead to unnecessary bulk of data and cause problems. In this paper, we consider the problem of recognizing objects in realistic senses. We first use Euclidean distance clustering method to segment objects in realistic scenes. Then we use a deep learning network structure to directly extract features of the point cloud data to recognize the objects. Theoretically, this network structure shows strong performance. In experiment, there is an accuracy rate of 98.8% on the training set, and the accuracy rate in the experimental test set can reach 89.7%. The experimental results show that the network structure in this paper can accurately identify and classify point cloud objects in realistic scenes and maintain a certain accuracy when the number of point clouds is small, which is very robust.

Multimodal medical datasets with incomplete observations present a barrier to large-scale neuroscience studies. Several works based on Generative Adversarial Networks (GAN) have been recently proposed to predict a set of medical images from a single modality (e.g., FLAIR MRI from T1 MRI). However, such frameworks are primarily designed to operate on images, limiting their generalizability to non-Euclidean geometric data such as brain graphs. While a growing number of connectomic studies has demonstrated the promise of including brain graphs for diagnosing neurological disorders, no geometric deep learning work was designed for multiple target brain graphs prediction from a source brain graph. Despite the momentum the field of graph generation has gained in the last two years, existing works have two critical drawbacks. First, the bulk of such works aims to learn one model for each target domain to generate from a source domain. Thus, they have a limited scalability in jointly predicting multiple target domains. Second, they merely consider the global topological scale of a graph (i.e., graph connectivity structure) and overlook the local topology at the node scale of a graph (e.g., how central a node is in the graph). To meet these challenges, we introduce MultiGraphGAN architecture, which not only predicts multiple brain graphs from a single brain graph but also preserves the topological structure of each target graph to predict. Its three core contributions lie in: (i) designing a graph adversarial auto-encoder for jointly predicting brain graphs from a single one, (ii) handling the mode collapse problem of GAN by clustering the encoded source graphs and proposing a cluster-specific decoder, (iii) introducing a topological loss to force the reconstruction of topologically sound target brain graphs. Our MultiGraphGAN significantly outperformed its variants thereby showing its great potential in multi-view brain graph generation from a single graph. Our code is available at https://github.com/basiralab/MultiGraphGAN.