Search

Real-time semantic segmentation, as a key technique for scene understanding, has been an important research topic in the field of computer vision in recent years. However, existing models are unable to achieve good segmentation accuracy on mobile devices due to their huge computational overhead, which makes it difficult to meet actual industrial requirements. To address the problems faced by current semantic segmentation tasks, this paper proposes a lightweight bilateral asymmetric residual network (LBARNet) for real-time semantic segmentation. First, we propose the bilateral asymmetric residual (BAR) module. This module learns multi-scale feature representations with strong semantic information at different stages of the semantic information extraction branch, thus improving pixel classification performance. Secondly, the spatial information extraction (SIE) module is constructed in the spatial detail extraction branch to capture multi-level local features of the shallow network to compensate for the lost geometric information in the downsampling stage. At the same time, we design the attention mechanism perception (AMP) module in the jump connection part to enhance the contextual representation. Finally, we design the dual branch feature fusion (DBF) module to exploit the correspondence between higher-order features and lower-order features to fuse spatial and semantic information appropriately. The experimental results show that LBARNet, without any pre-training and pre-processing and using only 0.6M parameters, achieves 70.8% mloU and 67.2% mloU on the Cityscapes dataset and Camvid dataset, respectively. LBARNet maintain a high segmentation accuracy while using a smaller number of parameters compared to most existing state-of-the-art models. [Display omitted] • This paper proposes a Bilateral Asymmetric Residual (BAR) module, a Spatial Information Extraction (SIE) module, an Attention Mechanism Perception (AMP) module and a Dual Branch Feature Fusion (DBF) module. • A Lightweight Bilateral Asymmetric Residual Network (LBARNet) for real-time image semantic segmentation is proposed in this article. • The experimental results show that LBARNet achieves 70.8% and 67.2% segmentation accuracy on two challenging datasets (Cityscapes and CamVid) using only 0.6M parameters. [ABSTRACT FROM AUTHOR]

Open-set semantic segmentation remains yet a challenging task, not only due to the inherent challenges of pixel-wise classification but also the precise segmentation of categories not seen during training. The pursuit of that task is rapidly growing in the Computer Vision community, urging the need to organize the literature. In this paper, we extend our previous work by conducting a more comprehensive systematic mapping of the open-set segmentation literature between January 2001 and January 2023 and proposing a novel taxonomy. Our goal is to provide a broad understanding of current trends for the open-set semantic segmentation (OSS) task defined by existing approaches that may influence future methods. By characterizing methodologies in terms of open-set identification strategies, data inputs, and other relevant aspects, we present a structured view of how researchers are advancing in the field of open-set semantic segmentation. To the best of the authors' knowledge, this is the first systematic review of OSS methods. Moreover, we apply the proposed taxonomy to selected methods for open-set recognition, outlining important similarities and differences of such a closely related field. [Display omitted] • Systematic review of papers related to open-set semantic segmentation for the past 20 years. • The proposed taxonomy aims to organize the literature on open-set segmentation. • Seminal papers on open-set recognition are classified under the proposed taxonomy. • Applications like autonomous driving and remote sensing were found to commonly resort to the open set strategy. • Methods tackling open-world are becoming more commom; [ABSTRACT FROM AUTHOR]

Low-light images often suffer from severe visual degradation, affecting both human perception and high-level computer vision tasks. Most existing methods process images in the spatial domain, making it challenging to simultaneously improve brightness while suppressing noise. In this paper, we present a novel perspective to enhance images based on frequency domain characteristics. Specifically, we reveal that the low-frequency components are closely related to luminance and color, whereas the high-frequency components are not. Based on this observation, we propose the Frequency-aware Network (FaNet) for low-light image enhancement. By selectively adjusting low-frequency components, FaNet preserves more high-frequency details while achieving low-light image enhancement. Additionally, we employ a multi-scale framework and selective fusion for effective feature learning and image reconstruction. Experimental results demonstrate the superiority of the proposed method. [Display omitted] • We reveal that the luminance is closely related to low-frequency components. • We design a frequency-aware network to utilize frequency domain features. • A multi-scale framework and selective fusion is proposed for feature learning. [ABSTRACT FROM AUTHOR]

Learning 3D mesh representation is necessary for many computer vision and graphic tasks. Recently, some works have studied convolution methods for directly processing input meshes. However, these methods are usually weak in extracting local geometry information because of the disadvantages such as isotropic filter, neglect of mesh topology, and a small convolution field. In this paper, we introduce a local geometry-perceptive mesh convolution, which pays attention to mesh irregular structures for efficiently capturing geometry features in a multi-ring receptive field. Specifically, we define each template node's dynamic neighbor-attention weights used in multiple attention aggregation operations for obtaining local mesh change information of different vertices in the multi-ring field. After each aggregation, a shared anisotropic filter maps the catenation of each new vertex and its neighbors for extracting geometry features of the current ring. Then, complete local geometry features of each vertex in its large local field are obtained by summing the mapped results of each aggregation. Moreover, the position features of each vertex are added to its local geometry features to get the final representation vector of the vertex. We demonstrate the proposed mesh convolution method's strong ability in modeling 3D mesh shapes. [Display omitted] • Utilizing local irregular mesh structures, dynamic neighbor-attention weights of template's nodes perceive local mesh change information of different vertices. • Multiple attention aggregations and shared anisotropic filters help extract geometry features of large local receptive fields with multi-ring. • The proposed method determines our deep 3D morphable models (3DMMs) have smaller sizes than previous models and achieve state-of-the-art performances. [ABSTRACT FROM AUTHOR]

Search and retrieval remains a major research topic in several domains, including computer graphics, computer vision, engineering design, etc. A search engine requires primarily an input search query and a database of items to search from. In engineering, which is the primary context of this paper, the database consists of 3D CAD models, such as washers, pistons, connecting rods, etc. A query from a user is typically in the form of a sketch, which attempts to capture the details of a 3D model. However, sketches have certain typical defects such as gaps, over-drawn portions (multi-strokes), etc. Since the retrieved results are only as good as the input query, sketches need cleaning-up and enhancement for better retrieval results. In this paper, a deep learning approach is proposed to improve or clean the query sketches. Initially, sketches from various categories are analysed in order to understand the many possible defects that may occur. A dataset of cleaned-up or enhanced query sketches is then created based on an understanding of these defects. Consequently, an end-to-end training of a deep neural network is carried out in order to provide a mapping between the defective and the clean sketches. This network takes the defective query sketch as the input and generates a clean or an enhanced query sketch. Qualitative and quantitative comparisons of the proposed approach with other state-of-the-art techniques show that the proposed approach is effective. The results of the search engine are reported using both the defective and enhanced query sketches, and it is shown that using the enhanced query sketches from the developed approach yields improved search results. [Display omitted] • The first learning-based strategy to clean rough query sketches of 3D CAD models • Introduces SketchCleanNet — an end-to-end image translation scheme • SketchCleanNet aims to understand the mapping between rough sketches and clean query images • A novel scheme to calculate the loss is introduced • Dataset Contribution: The resulting enhanced query sketch dataset is made available publicly. • This paper will significantly contribute to the research community and give researchers opportunities to develop new algorithms for search and retrieval of 3D mechanical components. [ABSTRACT FROM AUTHOR]

The research topic of estimating hand pose from the images of hand-object interaction has the potential for replicating natural hand behavior in many practical applications of virtual reality and robotics. However, the intricacy of hand-object interaction combined with mutual occlusion, and the need for physical plausibility, brings many challenges to the problem. This paper provides a comprehensive survey of the state-of-the-art deep learning-based approaches for estimating hand pose (joint and shape) in the context of hand-object interaction. We discuss various deep learning-based approaches to image-based hand tracking, including hand joint and shape estimation. In addition, we review the hand-object interaction dataset benchmarks that are well-utilized in hand joint and shape estimation methods. Deep learning has emerged as a powerful technique for solving many problems including hand pose estimation. While we cover extensive research in the field, we discuss the remaining challenges leading to future research directions. [Display omitted] • Deep learning is effectively used for estimating hand pose from images. • The correlation between a hand and an object helps in estimating hand-object pose. • Hand model helps estimate hand shape, but it restricts within the model's prior. • Implicit function methods have emerged in hand-object pose estimation. [ABSTRACT FROM AUTHOR]

Arbitrary style transfer has broad application prospects and important research values, and it is a research hotspot in the field of computer vision. Many studies have shown that arbitrary style transfer achieves remarkable success. However, existing methods may produce artifacts and sometimes lead to distortion of the content structure. To overcome this limitation, this paper proposes a novel Attention-wise and Covariance-Matching Module (ACMM) that preserves the content structure well without unpleasant artifacts. First, our method uses global attention covariance matching to match the global information of style features with content features, thereby obtaining pleasing stylized images. Then, to enable the model to better match the global statistics, a histogram loss is introduced to improve the saturation and stability of the resulting color. Our method can preserve the content structure; so that appearance transfer can be achieved with simple adjustments to the model. The effectiveness of the proposed method is demonstrated through qualitative and quantitative experiments with comparisons with some state-of-the-art arbitrary style transfer methods. [Display omitted] • The Attention-wise and Covariance-Matching Module can preserve content structure well without unpleasant artifacts. • The histogram loss can improve the color saturation of the generated results and enhances the stability of the results. • The proposed framework can be extended to achieve photorealistic style transfer. • The visual performance of the proposed framework is comparable to other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Multi-View Stereo (MVS) is a crucial technique for reconstructing the geometric structure of a scene, given the known camera parameters. Previous deep learning-based MVS methods have mainly focused on improving the reconstruction quality but overlooked the running efficiency during the actual algorithm deployment. For example, deformable convolutions have been introduced to improve the accuracy of the reconstruction results further, however, its inability for parallel optimization caused low inference speed. In this paper, we propose AdaptMVSNet which is device-friendly and reconstruction-efficient, while preserving the original results. To this end, adaptive convolution is introduced to significantly improve the efficiency in speed and metrics compared to current methods. In addition, an attention fusion module is proposed to blend features from adaptive convolution and the feature pyramid network. Our experiments demonstrate that our proposed approach achieves state-of-the-art performance and is almost 2 × faster than the recent fastest MVS method. We will release our source code. [Display omitted] [ABSTRACT FROM AUTHOR]

In this paper, we address the problem of capturing both the shape and the pose of a human character using a single depth sensor. Some previous works proposed to fit a parametric generic human template into the depth image, while others developed deep learning (DL) approaches to find the correspondence between depth pixels and vertices of the template. We designed a hybrid approach, combining the advantages of both methods, and conducted extensive experiments on the SURREAL Varol et al. (2017), DFAUST datasets Bogo etal (2017) and a subset of AMASS Mahmood et al. (2019). Results show that this hybrid approach enables us to enhance pose and shape estimation compared to using DL or model fitting separately. We also evaluated the ability of the DL-based dense correspondence method to segment also the background — not only the body parts. We also evaluated 4 different methods to perform the model fitting based on a dense correspondence, where the number of available 3D points differs from the number of corresponding template vertices. These two results enabled us to better understand how to combine DL and model fitting, and the potential limits of this approach to deal with real-depth images. Future works could explore the potential of taking temporal information into account, which has proven to increase the accuracy of pose and shape reconstruction based on a unique depth or RGB image. [Display omitted] • Dense correspondence and model fitting enhance 3D shape and pose reconstruction. • Bodypart and background color embedding enables character segmentation in depth map. • Averaging 3D candidate points before model fitting offers the best reconstruction. [ABSTRACT FROM AUTHOR]

Low-light image enhancement is crucial in applications such as traffic safety and medical imaging. Besides having characteristics like low luminance and poor visibility, low-light images are inevitably affected by noise. Noise not only covers image details, introduces artifacts, and decreases image quality, but also interferes with downstream computer vision tasks like object detection, image segmentation, and object tracking. Previous methods in image enhancement often overlook noise handling or fail to accurately suppress noise in adaptive denoising processes, resulting in more severe noise artifacts in the enhanced images. To effectively suppress noise, this paper proposes a Channel Self-Attention Based Low-Light Image Enhancement Network (CAENet), which leverages Transformers and CNNs to model long-range and short-range pixel dependencies, extract global and local features, and construct a Noise Suppression Transformer Block that adaptively suppresses noise regions guided by signal-to-noise ratio priors and attention maps. After adaptive noise suppression, the resulting images exhibit fewer noise artifacts and improved details. The experimental results show that the network in this paper outperforms other state-of-the-art methods overall on five representative paired datasets as well as six unpaired datasets, improving the image quality while effectively suppressing the noise. [Display omitted] • To address the problem of diverse noise types and complex interference levels in low-light image enhancement, we propose an end-to-end network CAENet. • Since the severity of noise varies in regions with different luminance, we designed the NSTB based on channel self-attention. • Aiming at the different degree of interference of noise to the whole and details of the image in different sized regions, we design global and local branches. [ABSTRACT FROM AUTHOR]

Scene understanding has made tremendous progress over the past few years, as data acquisition systems are now providing an increasing amount of data of various modalities (point cloud, depth, RGB...). However, this improvement comes at a large cost on computation resources and data annotation requirements. To analyze geometric information and images jointly, many approaches rely on both a 2D loss and 3D loss, requiring not only 2D per pixel-labels but also 3D per-point labels. However, obtaining a 3D groundtruth is challenging, time-consuming and error-prone. In this paper, we show that image segmentation can benefit from 3D geometric information without requiring a 3D groundtruth, by training the geometric feature extraction and the 2D segmentation network jointly, in an end-to-end fashion, using only the 2D segmentation loss. Our method starts by extracting a map of 3D features directly from a provided point cloud by using a lightweight 3D neural network. The 3D feature map, merged with the RGB image, is then used as an input to a classical image segmentation network. Our method can be applied to many 2D segmentation networks, improving significantly their performance with only a marginal network weight increase and light input dataset requirements, since no 3D groundtruth is required. • Geometry. • Computer vision. • 2D segmentation. • Neural network. • multi-modalities. [ABSTRACT FROM AUTHOR]

Reconstructing the shape of hand-held objects from single-view color images is a long-standing problem in computer vision and computer graphics. The task is complicated by the ill-posed nature of single-view reconstruction, as well as potential occlusions due to both the hand and the object. Previous works mostly handled the problem by utilizing known object templates as priors to reduce the complexity. In contrast, our paper proposes a novel approach without knowing the object templates beforehand but by exploiting prior knowledge of contacts in hand-object interactions to train an attention-based network that can perform precise hand-held object reconstructions with only a single forward pass in inference. The network we propose encodes visual features together with contact features using a multi-head attention module as a way to condition the training of a neural field representation. This neural field representation outputs a Signed Distance Field representing the reconstructed object and extensive experiments on three well-known datasets demonstrate that our method achieves superior reconstruction results even under severe occlusion compared to the state-of-the-art techniques. • Conditioning on contact, we reconstruct the hand-held object from single-view images. • We use an end-to-end attention-based network to better encode contact priors. • We achieve state-of-the-art results on Obman, HO3D, and MOW datasets. [ABSTRACT FROM AUTHOR]

Semantic segmentation is a difficult task in computer vision that have applications in many scenarios, often as a preprocessing step for a tool. Current solutions are based on Deep Neural Networks, which often require a large amount of data for learning a task. Aiming to alleviate the strenuous data-collecting/annotating labor, research fields have emerged in recent years. One of them is Meta-Learning, which tries to improve the generability of models to learn in a restricted amount of data. In this work, we extend a previous paper conducting a more extensive overview of the still under-explored problem of Few-Shot Weakly-supervised Semantic Segmentation. We refined the previous taxonomy and included the review of additional methods, including Few-Shot Segmentation methods that could be adapted to the weak supervision. The goal is to provide a simple organization of literature and highlight aspects observed in the current moment, and be a starting point to foment research on this problem with applications in areas like medical imaging, remote sensing, video segmentation, and others. [Display omitted] • A review of Non-Meta-Learning methods for weakly-supervised segmentation. • A review of Meta-Learning approaches for few-shot segmentation and classification. • A proposed taxonomy to Meta-Learning methods for few-shot semantic segmentation. • Limitations and future directions for few-shot weak-supervised segmentation. [ABSTRACT FROM AUTHOR]

2D poses are insufficient to estimate human activities under the geometric ambiguities, occlusion, diverse appearance, and viewpoints inside the image. As a result, 3D posing has emerged as an attractive research area in recent decades due to its ability to distinguish human activities effectively. Human skeletal 3D posture is the foundation for 3D avatar creation, the construction of human 3D meshes, recognition of human actions or activities, augmented reality, etc. Therefore, estimating the correct human 3D skeleton or pose with minimum errors is an essential task. This research paper proposes a depth predictor module; an innovative and creative approach of depth prediction for evaluating the human 3D pose from a single 2D image. However, no predicted 3D pose is entirely free from errors. Therefore, it needs to reduce the inaccuracy after predicting the 3D pose. For this purpose, an enhancement approach, pose alignment , is used to reduce the positioning inaccuracies of the anticipated 3D pose. The specialty of the proposed model is that it does not require any training dataset, and it effectively estimates 3D poses under any scale variations. This proposed method is evaluated on different datasets: Human3.6M, HumanEva-I, NTU RGB + D, MPII, UAV-human and the Articulated Free Fall dataset. In terms of errors, the proposed approach achieves the minimum 3D pose error compared to the existing 3D pose estimation techniques. The proposed depth prediction approach reduces the average MPJPE value to 48.30, and the pose alignment module further reduces it to 40.61. In addition, the proposed model takes an inference time of 1.806 s per frame for a single person. [Display omitted] • We introduced a novel concept for human 3D pose estimation from a single RGB image. • The specialty of this framework is that there is no need for any training dataset. • This is an algorithm-based system to estimate the 3D pose of the human body. • We introduced a Depth Predictor module, our novel concept to predict depth. • Proposed a pose alignment technique to enhance the accuracy of the predicted 3D pose. • This approach is novel and very interesting in the field of computer vision. [ABSTRACT FROM AUTHOR]

Outdoor point-cloud object localization is an essential processing step for urban scene analysis and modeling in numerous applications, especially in land surveying and site analysis. Given the increasingly use of Terrestrial Laser Scanning (TLS) and LiDAR 3D data acquisition, numerous annotated point-cloud datasets are available and can be used to evaluate computer vision and machine learning based algorithms. Nevertheless, current point-cloud datasets mainly focus on object detection and classification in autonomous driving or urban planning type of applications, and share redundant object classes, e.g., trees, vehicles and pedestrians, which limit their usefulness for land surveying and site analysis. This paper introduces a novel 3D benchmark dataset LiSurveying, which is a large-scale point-cloud dataset with over a billion points and uncommon urban object categories in complex outdoor environments. Our dataset incorporates more urban object classes than existing datasets. Its instances have diverse point densities, shapes and dimensions, which also impose a challenge for point-cloud detection and classification algorithms. We conducted baselines experiments for point-cloud classification using machine learning classifiers and deep learning methods on different subsets of the LiSurveying dataset, and we are able to demonstrate that the various types of object classes, number of instances per class, distribution of object points, and variety of complex scenes, make this LiSurveying benchmark dataset suitable for evaluating 3D point-cloud classification, semantic segmentation, and object detection algorithms. [Display omitted] • Large-scale (point-wise and 3D bounding boxes) TLS-LiDAR dataset for Land Surveying. • 3D Urban benchmark for object classification y detection. • Learning-based methods applied on complex urban environments. • Comparative performance of existing algorithms for LiDAR classification. [ABSTRACT FROM AUTHOR]

Deep neural networks exhibit extreme superiority in natural language processing and computer vision. With the fast development of 3D sensing technology, point cloud has been wildly used in robotics and autonomous driving. In recent years, several attempts have been made to utilize deep neural networks in the field of point cloud processing. However, most top performing network cannot provide the confidence of each prediction. In this paper, we implement Bayesian deep learning method in point cloud processing and propose ProbConv, a three-dimensional convolutional kernel with stochastic weights. Based on ProbConv, a Bayesian deep neural network named ProbNet is further designed to effectively accomplish classification and segmentation tasks on point cloud data. Since Bayesian neural network can naturally calculate the confidence of its prediction, ProbNet is able to provide the confidence of the segmentation result. The experimental results on ModelNet40, ShapeNet and S3DIS demonstrate that ProbNet improves accuracy in both object classification and semantic segmentation tasks. The confidence provided by the ProbNet is reliable to reflect the accuracy of its prediction. [Display omitted] • Bayesian deep learning and pyramid-divided respective field is implemented on ProbConv. • ProbNet is based on the proposed ProbConv operator to provide confidence of its prediction. • Experiment result shows that ProbNet improves accuracy while providing reliable confidence. [ABSTRACT FROM AUTHOR]

Indoor navigation has remained an active research area for the last decade. Unlike outdoor environments, indoor environments have additional challenges, such as weak signals, low light, and complex scenarios. Different technologies are used for indoor navigation, including WiFi, Bluetooth, inertial sensors, and computer cameras. Vision-based methods have great potentials for indoor navigation as they fulfill most of the general requirements such as minimal cost, ease of use, ease of implementation, and realism. Therefore, researchers have successively proposed different novel vision-based approaches for indoor navigation. Unfortunately, there is no standard review article (except a few general reviews) that covers the current trends and draws a pipeline for future research. In this paper, we reviewed the current state-of-the-art vision-based indoor navigation methods. We followed the systematic literature review (SLR) methodology for article searching, selection, and quality assessments. In total, we selected 68 articles after final selection using SLR. We classified these articles into different categories. Each article is briefly studied for information extraction, including key idea, category of the article, evaluation criterion, and its strengths and weaknesses. We also highlighted several interesting future directions. This study will help new researchers to grasp the research challenge as well as present the results of their research in the field. It will also help the community to find a suitable indoor navigation system according to users' requirements. [Display omitted] • Executed the first SLR study on vision-based methods for indoor navigation. • As a result, 68 studies were found, analyzed, and presented in six categories. • Each article is analyzed for data extraction and its quality assessment. • We extracted key idea, pros and con, and analysis metrics from each article. • A number of interested future direction have been drawn. [ABSTRACT FROM AUTHOR]

Obtaining good feature descriptors for 3D local shape under different conditions, such as noise, clutter, occlusion, and limited overlap, is still a challenging task in computer vision. In this paper, we present a variant of spin-image called dual spin-image (Dual-SI) for 3D local shape description that encodes bi-directional information between an oriented point and its surrounding neighbors within a given radius. By accumulating the parameters that encode the 3D point information into a 2D space bidirectionally, we can generate two different signatures. These signatures are then concatenated to form the Dual-SI feature descriptor. Finally, we propose a multi-scale radius approach to address the problem of occlusion and make use of a weighted kernel to address the noise problem. We tested our Dual-SI feature descriptor on popular datasets addressing 3D object registration, 3D object recognition, and shape retrieval scenarios. We also conduct experiments for 3D point cloud registration to further evaluate the effectiveness of our method. Consensus experimental results show that our Dual-SI achieves outstanding performance on datasets with various nuisances and application contexts. [Display omitted] • A Dual-SI variant of the spin-image descriptor that uses bidirectional encoding. • Dual-SI uses multi-scale radii to address occlusion problems. • Dual-SI uses weighted kernel to reduce the noise effect on scanned data. • Dual-SI achieves balanced performance in different experimental conditions. • Dual-SI offers sufficient consistent feature matches between point clouds. [ABSTRACT FROM AUTHOR]

3D reconstruction from a single image is one of the core computer vision problems. Thanks to the development of deep learning, 3D reconstruction of a single image has demonstrated impressive progress in recent years. Existing researches use Chamfer distance as a loss function to guard the training of the neural network. However, the Chamfer loss will give equal weights to all points inside the 3D point clouds. It tends to sacrifice fine-grained and thin structures to avoid incurring a high loss, which will lead to visually unsatisfactory results. This paper proposes a framework that can recover a detailed three-dimensional point cloud from a single image by focusing more on boundaries (edge and corner points). Experimental results demonstrate that the proposed method outperforms existing techniques significantly, both qualitatively and quantitatively, and has fewer training parameters. [Display omitted] • Reconstructing the full 3D shape of an object from a single image. • Using a differentiable point cloud projection module to get the projected images. • Increasing visual reconstruction quality. [ABSTRACT FROM AUTHOR]

Point set registration has been actively studied in computer vision and graphics. Optimization algorithms are at the core of solving registration problems. Traditional optimization approaches are mainly based on the gradient of objective functions. The derivation of objective functions makes it challenging to find optimal solutions for complex optimization models, especially for those applications where accuracy is critical. Learning-based optimization is a novel approach to address this problem, which learns the gradient direction from datasets. However, many learning-based optimization algorithms learn gradient directions via a single feature extracted from the dataset, which will cause the updating direction to be vulnerable to perturbations around the data, thus falling into a bad stationary point. This paper proposes the General Discriminative Optimization (GDO) method that updates a gradient path automatically through the trade-off among contributions of different features on updating gradients. We illustrate the benefits of GDO with tasks of 3D point set registrations and show that GDO outperforms the state-of-the-art registration methods in terms of accuracy and robustness to perturbations. [Display omitted] • Cast the point sets registration as a learning-based optimization problem. • Utilize features of point sets to update gradients without the Hessian matrix. • Collaborate the different features to reduce the influence of perturbations. • Outperform the other advanced registration methods on accuracy and robustness. [ABSTRACT FROM AUTHOR]

• Point inception based deep neural network for segmentation of the 3D point clouds. • Extract the local features using inception module to improve the performance. • Use GAP over other pooling methods to avoid overfitting. • Exhaustive evaluation and comparative analysis with the existing methods on two state-of-the-art datasets, namely, ShapeNet and PartNet. • Additional architecture and the ablation study demonstrate the effectiveness of proposed model. [Display omitted] Point clouds, being the simple and compact representation of surface geometry of 3D objects, have gained increasing popularity with the evolution of deep learning networks for classification and segmentation tasks. Unlike human, teaching the machine to analyze the segments of an object is a challenging task and quite essential in various machine vision applications. In this paper, we address the problem of segmentation and labelling of the 3D point clouds by proposing a inception based deep network architecture called PIG-Net, that effectively characterizes the local and global geometric details of the point clouds. In PIG-Net, the local features are extracted from the transformed input points using the proposed inception layers and then aligned by feature transform. These local features are aggregated using the global average pooling layer to obtain the global features. Finally, feed the concatenated local and global features to the convolution layers for segmenting the 3D point clouds. We perform an exhaustive experimental analysis of the PIG-Net architecture on two state-of-the-art datasets, namely, ShapeNet [1] and PartNet [2]. We evaluate the effectiveness of our network by performing ablation study. [ABSTRACT FROM AUTHOR]

• Computer graphics. • Computer vision. • Deep learning. • 3D reconstruction. • Virtual human digitization. • Generative adversarial network. [Display omitted] With the development of AR/VR technologies, a reliable and straightforward way to digitize a three-dimensional human body is in high demand. Most existing methods use complex equipment and sophisticated algorithms, but this is impractical for everyday users. In this paper, we propose a pipeline that reconstructs a 3D human shape avatar from a single image. Our approach simultaneously reconstructs the three-dimensional human geometry and whole body texture map with only a single RGB image as input. We first segment the human body parts from the image and then obtain an initial body geometry by fitting the segment to a parametric model. Next, we warp the initial geometry to the final shape by utilizing a silhouette-based dense correspondence. Finally, to infer invisible back texture from a frontal image, we propose a network called InferGAN. Based on human semantic information, we also propose a method to handle partial occlusion by reconstructing the occluded body parts separately. Comprehensive experiments demonstrate that our solution is robust and effective on both public and our own datasets. Our human avatars can be easily rigged and animated using MoCap data. We have developed a mobile application that demonstrates this capability for AR applications. [ABSTRACT FROM AUTHOR]

Although the concept of Augmented Reality (AR) has already been proposed more than 20 years ago, most AR-applications are still limited to simple visualization of virtual objects onto spatially limited scenes. Ideas such as the “augmented reality manual”, showing step-by-step instructions to a user wearing a Head-Mounted Display (HMD), have been implemented, but the developed systems did not pass the barrier of demonstration prototypes. One major reason, beside remaining ergonomic and hardware limitations, consists of the large effort required for creating the content of such virtual instructions and for building models allowing accurate tracking. In this paper, we introduce the concept of Cognitive Augmented Reality, which radically revises existing approaches on AR-based assistance systems and proposes a fundamentally new paradigm exploiting prior visual observation and learning of a complete manipulative workflow. More precisely, we present in this paper a complete approach for creating augmented reality content for procedural tasks from video examples, and give then details about the presentation of such content at runtime. We show that this new approach, in spite of its very challenging aspects, is scalable, and valuable from a practicable point of view. [ABSTRACT FROM AUTHOR]

• Publicly available, fully-annotated simulated cryo-electron tomogram dataset. • Benchmark of six learning-based methods against template matching baseline. • Strong correlation between localization/classification performance and molecular weight. Cryo-electron tomography (cryo-ET) is an imaging technique that allows us to three-dimensionally visualize both the structural details of macro-molecular assemblies under near-native conditions and its cellular context. Electrons strongly interact with biological samples, limiting electron dose. The latter limits the signal-to-noise ratio and hence resolution of an individual tomogram to about 50 (5 nm). Biological molecules can be obtained by averaging volumes, each depicting copies of the molecule, allowing for resolutions beyond 4 (0.4 nm). To this end, the ability to localize and classify components is crucial, but challenging due to the low signal-to-noise ratio. Computational innovation is key to mine biological information from cryo-electron tomography. To promote such innovation, we provide a novel simulated dataset to benchmark different methods of localization and classification of biological macromolecules in cryo-electron tomograms. Our publicly available dataset contains ten tomographic reconstructions of simulated cell-like volumes. Each volume contains twelve different types of complexes, varying in size, function and structure. In this paper, we have evaluated seven different methods of finding and classifying proteins. Six research groups present results obtained with learning-based methods and trained on the simulated dataset, as well as a baseline template matching, a traditional method widely used in cryo-ET research. We find that method performance correlates with particle size, especially noticeable for template matching which performance degrades rapidly as the size decreases. We learn that neural networks can achieve significantly better localization and classification performance, in particular convolutional networks with focus on high-resolution details such as those based on U-Net architecture. [ABSTRACT FROM AUTHOR]

• A new human pose regression method from RGB images. • The proposed soft-argmax operation. • Contextual aggregation for refining predictions. • State-of-the-art results on important 2D pose estimation benchmarks among regression methods. In this paper, we tackle the problem of human pose estimation from still images, which is a very active topic, specially due to its several applications, from image annotation to human-machine interface. We use the soft-argmax function to convert feature maps directly to body joint coordinates, resulting in a fully differentiable framework. Our method is able to learn heat maps representations indirectly, without additional steps of artificial ground truth generation. Consequently, contextual information can be included to the pose predictions in a seamless way. We evaluated our method on two challenging datasets, the Leeds Sports Poses (LSP) and the MPII Human Pose datasets, reaching the best performance among all the existing regression methods. Source code available at: https://github.com/dluvizon/pose-regression. [ABSTRACT FROM AUTHOR]

• A technique that performs semantic modeling and tracking of primitives on sparse point clouds on desktop and mobile devices. • The method allows the modeling and tracking of planes, spheres, and cylinders. • A dataset with sparse point clouds of primitives. • A guideline to evaluate computer vision techniques on mobile devices. Improvements on mobile devices allowed tracking applications to be executed on such platforms. However, there still remain several challenges in the field of mobile tracking, such as the extraction of high-level semantic information from point clouds. This task is more challenging when using monocular visual SLAM systems that output noisy sparse data. In this paper, we propose a primitive modeling method using both geometric and statistical analyses for sparse point clouds that can be executed on mobile devices. The main idea is to use the incremental mapping process of SLAM systems for analyzing the geometric relationship between the point cloud and the estimated shapes over time and selecting only reliably-modeled shapes. Besides that, a statistical evaluation that assesses if the modeling is random is incorporated to filter wrongly-detected primitives in unstable estimations. Our evaluation indicates that the proposed method was able to improve both precision and consistency of correct detection when compared with existing techniques. The mobile version execution is 8.5 to 9.9 times slower in comparison with the desktop implementation. However, it uses up to 30.5% of CPU load, which allows it to run on a separate thread, in parallel with the visual SLAM technique. Additional evaluations show that CPU load, energy consumption and RAM memory usage were not a concern when running our method on mobile devices. [ABSTRACT FROM AUTHOR]

Highlights • We present a geometric framework for 3D shape retrieval using geodesic moments. • We propose an unsupervised approach for learning shape descriptors using sparse autoencoders. • We demonstrate the superior performance of our approach on standard shape benchmarks. Graphical abstract Abstract Shape representations provide compact, parsimonious shape descriptions that are often used in object recognition and retrieval tasks. In light of the increased processing power of graphics cards and the availability of large-scale datasets, deep neural networks have shown a remarkable performance in numerous computer vision and geometry processing applications. In this paper, we present a deep learning framework for unsupervised 3D shape retrieval with geodesic moments. The proposed method learns deep shape representations using stacked sparse autoencoders in an unsupervised manner. Such discriminative shape descriptors can then be used to compute pairwise dissimilarities between shapes in a dataset, and to find the retrieved set of the most relevant shapes to a given shape query. Experimental evaluation on four standard 3D shape benchmarks demonstrate the competitive performance of our approach, showing that it leads to improved retrieval results in comparison with state-of-the-art techniques. [ABSTRACT FROM AUTHOR]

Panoramic imaging has important implications in robotics, computer vision and virtual reality. This paper reviews representative work in the design and development of 2D/3D panoramic image capturing systems, the advancement of auto-calibration, registration and corresponding techniques, stereo vision, 3D reconstruction and image-based rendering. The paper discusses the above work within four parts of the panoramic imaging process: capturing system, image processing for panoramic imaging, image stitching and 3D reconstruction, image-based rending and visualisation. The design and development of a panoramic system pays careful attention to the following issues: image capturing, image registration, camera calibration, feature extraction, image understanding and image stitching. The objective of this review paper is to summarise and compare some of the methods in the various stages and identify research topics and applications, which are at the forefront of this exciting and challenge field. [Copyright &y& Elsevier]

• VCBM is the Eurographics Workshop on Visual Computing for Biology and Medicine. • VCBM addresses the state of the art in visual computing research with a strong focus on applications in biology and medicine. • The topics of VCBM include visualization, visual analytics, computer graphics, image processing, computer vision, human computer interfaces, within the context of biology and medicine. • This section welcomes significantly extended and revised papers from VCBM 2021 and completely new articles within the same scope. [ABSTRACT FROM AUTHOR]

We propose a novel fully convolutional network architecture for shapes, denoted by Shape Fully Convolutional Networks (SFCN) . 3D shapes are represented as graph structures in the SFCN architecture, based on novel graph convolution and pooling operations , which are similar to convolution and pooling operations used on images. Meanwhile, to build our SFCN architecture in the original image segmentation fully convolutional network (FCN) architecture, we also design and implement a generating operation with bridging function. This ensures that the convolution and pooling operation we have designed can be successfully applied in the original FCN architecture. In this paper, we also present a new shape segmentation approach based on SFCN. Furthermore, we allow more general and challenging input, such as mixed datasets of different categories of shapes which can prove the ability of our generalisation. In our approach, SFCNs are trained triangles-to-triangles by using three low-level geometric features as input. Finally, the feature voting-based multi-label graph cuts is adopted to optimise the segmentation results obtained by SFCN prediction. The experiment results show that our method can effectively learn and predict mixed shape datasets of either similar or different characteristics, and achieve excellent segmentation results. [ABSTRACT FROM AUTHOR]

Decolorization is widely used in monotonic displaying, black-and-white printing, single-channel image processing, and many computer vision applications. The conversion of a color image into a grayscale one suffers from data loss. This paper presents a novel parallel decolorization method which effectively preserves not only the spatially local color contrast but also the dominant non-local color contrast. A new multimodal contrast-preserving measure with a multimodal Gaussian distribution is proposed to relax the constraint of color contrast. The dominant non-local color pair set is constructed by taking advantage of a linear bounding volume hierarchy while the local color pair set is produced by removing duplicate instances of local color pairs. The whole pipeline design is highly parallel, allowing for a real-time GPU-based implementation. Experimental results and comparisons show that the proposed method can produce plausible decolorized images. A number of single-channel image processing applications, including edge detection, image segmentation, color augmentation, and image stylization, are demonstrated to verify the feasibility of the proposed decolorization method. [ABSTRACT FROM AUTHOR]

Visual object tracking is a challenging research task in computer vision community which has been intensively studied by researchers in the past decades. Among all of the existing methods, particle filter based methods have gained special attention due to its ability to handle highly nonlinear/non-Gaussian multi-modality models. This paper proposes a robust particle filter based tracking method based on the Langevin Monte Carlo sampling. The Langevin Monte Carlo sampling method leverages the gradient of the posterior probability distribution to draw new particles. Meanwhile, an auxiliary momentum variable is introduced to ensure that the proposed sample cannot be trapped in local mode of the posterior distribution. As for the likelihood model, we introduce the locality sensitive histogram based model to handle the severe appearance variations induced by illumination change, partial occlusion or other factors. We compare the proposed method with several popular tracking methods from qualitative and quantitative perspectives. The experimental results show that the proposed method outperforms its counterparts. [ABSTRACT FROM AUTHOR]

Mesh saliency was introduced and joined the community of computer graphics ten years ago, which can benefit various applications, for instance, mesh reduction, mesh segmentation, self-similarity matching, scan integration, volume rendering, 3D printing, etc. Before, saliency detection had been successfully applied to image processing and pattern recognition to study how the world is perceptually intelligible for robots. In contrast to color of images and coherence of videos, geometric signals are defined with two-dimensional manifolds whose discrete representation is irregular, leading differences to the nature and difficulties to the solution of mesh saliency. To tackle the challenge, the last decade has witnessed significant advances in mesh saliency detection. However, a survey of recent advances in mesh saliency detection as well as its applications does not yet exist to date. This paper provides a first and comprehensive reference source of shape context based mesh saliency for researchers from a wide range of domains, including but not limited to computer graphics and vision. It reviews main contributions, advantages, drawbacks, and applications of known mesh saliency detection methods and discusses current trends and outlook for future study. [ABSTRACT FROM AUTHOR]

Sketching is one of the simplest ways to visualize ideas. Its key advantage is its easy availability and accessibility, as it require the user to have neither deep knowledge of a particular drawing program nor any advanced drawing skills. In practice, however, all these skills become necessary to improve the visual fidelity of the resulting drawing. In this paper, we present ShipShape—a general beautification assistant that allows users to maintain the simplicity and speed of freehand sketching while still taking into account implicit geometric relations to automatically rectify the output image. In contrast to previous approaches ShipShape works with general Bézier curves, enables undo/redo operations, is scale independent, and is fully integrated into Adobe Illustrator. We show various results to demonstrate the capabilities of the proposed method. [ABSTRACT FROM AUTHOR]

Video segmentation is an important preprocessing step for many computer vision and graphics tasks. Its main goal is to group the voxels in the video volume with similar appearance and motion into spatio-temporally consistent supervoxels. In this paper, we formulate video segmentation as an L 0 gradient minimization problem, so that the spatio-temporal coherence can be effectively enforced through a gradient sparsity pursuit way. In our method, the appearance and motion descriptor space is first built for over-segmented image patches of each video frame. Then the L 0 gradient minimization is performed in the descriptor space, for both spatial and temporal dimensions. To solve the non-convex L 0 norm minimization problem, we extend the fused coordinate descent algorithm from 2D image grids to 3D video volume. We conduct quantitative evaluation of our method in a public video segmentation benchmark LIBSVX. The experimental results demonstrate our superior performance to state-of-the-arts in segmentation accuracy and undersegmentation error, and comparable performance in boundary recall and explained variation. [ABSTRACT FROM AUTHOR]

3D Heritage Online Presenter (3DHOP) is a framework for the creation of advanced web-based visual presentations of high-resolution 3D content. 3DHOP has been designed to cope with the specific needs of the Cultural Heritage (CH) field. By using multiresolution encoding, it is able to efficiently stream high-resolution 3D models (such as the sampled models usually employed in CH applications); it provides a series of ready-to-use templates and examples tailored for the presentation of CH artifacts; it interconnects the 3D visualization with the rest of the webpage DOM, making it possible to create integrated presentations schemes (3D + multimedia). In its design and development, we paid particular attention to three factors: easiness of use, smooth learning curve and performances. Thanks to its modular nature and a declarative-like setup, it is easy to learn, configure, and customize at different levels, depending on the programming skills of the user. This allows people with different background to always obtain the required power and flexibility from the framework. 3DHOP is written in JavaScript and it is based on the SpiderGL library, which employs the WebGL subset of HTML5, implementing plugin-free 3D rendering on many web browsers. In this paper we present the capabilities and characteristics of the 3DHOP framework, using different examples based on concrete projects. [ABSTRACT FROM AUTHOR]

We present a detailed exploration of VideoHandles , a novel interaction technique to support rapid review of wearable video camera data by re-performing gestures as a search query. The availability of wearable video capture devices has led to a significant increase in activity logging across a range of domains. However, searching through and reviewing footage for data curation can be a laborious and painstaking process. We showcase the use of gestures as search queries to support review and navigation of video data. By exploring example action-camera footage across a range of activities, we propose two video data navigation styles using gestures: prospective gesture tagging and retrospective gesture searching . This paper builds on our earlier VideoHandles work, reporting details of our interaction design and implementation, and presenting two additional evaluations. We demonstrate that VideoHandles is a viable interaction technique, returning promising results in its prototype form. [ABSTRACT FROM AUTHOR]

Abstract: Feature-preserving filtering is a fundamental tool in computer vision and graphics, which can smooth input signal while preserving its sharp features. Recently, a piecewise smooth model called L 0 gradient minimization, has been proposed for feature-preserving filtering. Through optimizing an energy function involving gradient sparsity prior, L 0 gradient minimization model has strong ability to keep sharp features. Meanwhile, due to the non-convex property of L 0 term, it is a challenge to solve the L 0 gradient minimization problem. The main contribution of this paper is a novel and efficient approximation algorithm for it. The energy function is optimized in a fused coordinate descent framework, where only one variable is optimized at a time, and the neighboring variables are fused together once their values are equal. We apply the L 0 gradient minimization in two applications: (i) edge-preserving image smoothing (ii) feature-preserving surface smoothing, and demonstrate its good performance. [Copyright &y& Elsevier]

Image, graphical abstract • VCBM is the Eurographics Workshop on Visual Computing for Biology and Medicine. • VCBM addresses the state of the art in visual computing research with a strong focus on applications in biology and medicine. • The topics of VCBM include visualization, visual analytics, computer graphics, image processing, computer vision, human computer interfaces. • This section contains significantly extended and revised papers from VCBM 2019 and completely new articles within the same scope. [ABSTRACT FROM AUTHOR]

Abstract: This paper presents a method to assist in the tedious task of reconstructing ceramic vessels from unearthed archeological shards (fragments) using 3D computer vision-enabling technologies. The method exploits the shards surface intrinsic differential geometry information (one of many possible tools) coupled with a series of generic models to produce a virtual reconstruction and rendition of what the original vessel may have looked like. Generic models are constructed based on a host of factors including expert historical knowledge of the period, provenance of the artifact, and site location. The generic models need not to be identical to the original vessel, but they must be within a geometric transformation of it in most of its parts. The method is suited for ceramic vessels with some relief (i.e., surface with molding, carving, or stamping), as this method exploits surface intrinsic features for alignment. The alignment of the shards against the generic model uses a novel set of 3D weighted curve moments. The transformation is computed from corresponding parabolic contours on the shard and the generic model. A distance error metric is used to access the accuracy of alignment of a fragment to a given generic vessel model. The method is also extendable to surface markings. For a vessel that has no relief, color information or surface breaks can be used for the alignment. The method is tested on a subset of 3D scanned Independence National Historical Park (INDE) ceramic artifacts and the generic models created by experts. This work is part of an ongoing research activity in computational archeology that exploits many different tools and features to help in the mending process. Only the use of surface differential geometry information is reported here. This aspect is complementary to various other tools reported elsewhere by us and others, such as surface breaks, texture, color, etc. [Copyright &y& Elsevier]

Abstract: This article is a revised version of an earlier work on Gravity-Aware Handheld Augmented Reality (AR) (Kurz and Benhimane, 2011 [1]), which investigates how different stages in handheld AR applications can benefit from knowing the direction of the gravity measured with inertial sensors. It presents approaches to improve the description and matching of feature points, detection and tracking of planar templates, and the visual quality of the rendering of virtual 3D objects by incorporating the gravity vector. In handheld AR, both the camera and the display are located in the user''s hand and therefore can be freely moved. The pose of the camera is generally determined with respect to piecewise planar objects that have a static and known orientation with respect to gravity. In the presence of (close to) vertical surfaces, we show how Gravity-Aligned Feature Descriptors (GAFDs) improve the initialization of tracking algorithms relying on feature point descriptor-based approaches in terms of quality and performance. For (close to) horizontal surfaces, we propose to use the gravity vector to rectify the camera image and detect and describe features in the rectified image. The resulting Gravity-Rectified Feature Descriptors (GREFDs) provide an improved precision–recall characteristic and enable faster initialization, in particular under steep viewing angles. Gravity-rectified camera images also allow for real-time 6 DoF pose estimation using an edge-based object detection algorithm handling only 4 DoF similarity transforms. Finally, the rendering of virtual 3D objects can be made more realistic and plausible by taking into account the orientation of the gravitational force in addition to the relative pose between the handheld device and a real object. In comparison to the original paper, this work provides a more elaborate evaluation of the presented algorithms. We propose a method enabling the evaluation of inertial-sensor aided visual tracking methods without real inertial sensor data. By synthesizing gravity measurements from ground truth camera poses, we benchmark our algorithms on a large existing dataset. Based on this approach, we also develop and evaluate a gravity-adaptive approach that performs image-rectification only when beneficial. [Copyright &y& Elsevier]

Abstract: Highly detailed natural scenes and objects tend to be perceived as being realistic, while repeated parts and patterns decrease their realism. To avoid scenes with noticeable repeated elements, we introduce the notion of ‘more of the same’, which focuses on the task of generating additional similar instances from a small set of exemplars. The small number of exemplars, as well as their diversity and detailed structural texture, makes it difficult to apply statistical methods, or other machine learning tools, and thus more specialized tools need to be used. In this paper, we focus on generating a rich variation of highly detailed realistic leaves from just a handful set of examples. The method that we present does use only minimal domain specific knowledge and requires only minimal user assistance applied on a single training leaf exemplar to extract and separate structural layers. The knowledge from one leaf is then transferred to the other exemplars by a novel color/spatial layer inducing algorithm. The premise of structural layering is that each set of layers is simple enough to be synthesized separately and then composed into a novel leaf structural texture. This composition also allows the synthesis of slightly modified layers from the set of examples, which can generate a large set of differently looking leaves. We demonstrate numerous results of realistically looking leaves produced by our method from a small set of leaves. [Copyright &y& Elsevier]

Abstract: Developing augmented reality (AR) applications for mobile devices and outdoor environments has historically required a number of technical trade-offs related to tracking. One approach is to rely on computer vision which provides very accurate tracking, but can be brittle, and limits the generality of the application. Another approach is to rely on sensor-based tracking which enables widespread use, but at the cost of generally poor tracking performance. In this paper we present and evaluate a new approach, which we call Indirect AR, that enables perfect alignment of virtual content in a much greater number of application scenarios. To achieve this improved performance we replace the live camera view used in video see through AR with a previously captured panoramic image. By doing this we improve the perceived quality of the tracking while still maintaining a similar overall experience. There are some limitations of this technique, however, related to the use of panoramas. We evaluate these boundaries conditions on both a performance and experiential basis through two user studies. The result of these studies indicates that users preferred Indirect AR over traditional AR in most conditions, and when conditions do degrade to the point the experience changes, Indirect AR can still be a very useful tool in many outdoor application scenarios. [Copyright &y& Elsevier]

Abstract: Accurate geometric properties estimation from discrete curves is an important problem in many application domains, such as computer vision, pattern recognition, image processing, and geometric modeling. In this paper, we propose a novel method for estimating the geometric properties from discrete curves based on derivative estimation. We develop derivative estimation by defining the derivative of a discrete function at a point, which will be called the discrete derivative. Similarly, the second and higher order discrete derivatives at that point are also defined, and their convergence is demonstrated by theory analysis. These definitions of the different order discrete derivatives provide a simple and reliable way to estimate the derivatives from discrete curves. Based on the discrete derivatives, classical differential geometry can be discretized, and the geometric properties are estimated from discrete curves by using differential geometry theory. The proposed method is independent of any analytic curve and estimates the geometric properties directly from discrete data points, which makes it robust to the geometric shapes of discrete curves. Another advantage of the proposed method is the robustness to noise because of the calculation characteristics of the discrete derivatives. The proposed method is evaluated and compared with other existing methods in the experiments with both synthetic and real discrete curves. The test results show that the proposed method has good performance, and is robust to noise and suitable for different curve shapes. [Copyright &y& Elsevier]

Abstract: In this paper we present a method for the calibration of multiple cameras based on the extraction and use of the physical characteristics of a one-dimensional invariant pattern which is defined by four collinear markers. The advantages of this kind of pattern stand out in two key steps of the calibration process. In the initial step of camera calibration methods, related to sample points capture, the proposed method takes advantage of using a new technique for the capture and recognition of a robust sample of projective invariant patterns, which allows to capture simultaneously more than one invariant pattern in the tracking area and recognize each pattern individually as well as each marker that composes them. This process is executed in real time while capturing our sample of calibration points in the cameras of our system. This new feature allows to capture a more numerous and robust set of sample points than other patterns used for multi-camera calibration methods. In the last step of the calibration process, related to camera parameters'' optimization, we explore the collinearity feature of the invariant pattern and add this feature in the camera parameters optimization model. This approach obtains better results in the computation of camera parameters. We present the results obtained with the calibration of two multi-camera systems using the proposed method and compare them with other methods from the literature. [Copyright &y& Elsevier]