Within (semi-)automated visual industrial inspection, learning-based approaches for assessing visual defects, including deep neural networks, enable the processing of otherwise small defect patterns in pixel size on high-resolution imagery. The emergence of these often rarely occurring defect patterns explains the general need for labeled data corpora. To alleviate this issue and advance the current state of the art in unsupervised visual inspection, this work proposes a DifferNet-based solution enhanced with attention modules: AttentDifferNet. It improves image-level detection and classification capabilities on three visual anomaly detection datasets for industrial inspection: InsPLAD-fault, MVTec AD, and Semiconductor Wafer. In comparison to the state of the art, AttentDifferNet achieves improved results, which are, in turn, highlighted throughout our quali-quantitative study. Our quantitative evaluation shows an average improvement - compared to DifferNet -of 1.77 ± 0.25 percentage points in overall AUROC considering all three datasets, reaching SOTA results in InsPLAD-fault, an industrial inspection in-the-wild dataset. As our variants to AttentDifferNet show great prospects in the context of currently investigated approaches, a baseline is formulated, emphasizing the importance of attention for industrial anomaly detection both in the wild and in controlled environments.

Volumetric radiance fields have been popular in reconstructing small-scale 3D scenes from multi-view images. With additional constraints such as person correspondences, reconstructing a large 3D scene with multiple persons becomes possible. However, existing methods fail for sparse input views or when person correspondences are unavailable. In such cases, the conventional depth image supervision may be insufficient because it only captures the relative position of each person with respect to the camera center. In this paper, we investigate an alternative approach by supervising the optimization framework with a dense pose prior that represents correspondences between the SMPL model and the input images. The core ideas of our approach consist in exploiting dense pose priors estimated from the input images to perform person segmentation and incorporating such priors into the learning of the radiance field. Our proposed dense pose supervision is view-independent, significantly speeding up computational time and improving 3D reconstruction accuracy , with less floaters and noise. We confirm the advantages of our proposed method with extensive evaluation in a subset of the publicly available CMU Panoptic dataset. When training with only five input views, our proposed method achieves an average improvement of 6.1% in PSNR , 3.5% in SSIM , 17.2% in LPIPS vgg , 19.3% in LPIPS alex , and 39.4% in training time.

The detection of objects considering a 6DoF pose is a common requirement to build virtual and augmented reality applications. It is usually a complex task which requires real-time processing and high precision results for adequate user experience. Recently, different deep learning techniques have been proposed to detect objects in 6DoF in RGB images. However, they rely on high complexity networks, requiring a computational power that prevents them from working on mobile devices. In this paper, we propose an approach to reduce the complexity of 6DoF detection networks while maintaining accuracy. We used Knowledge Distillation to teach portables Convolutional Neural Networks (CNN) to learn from a real-time 6DoF detection CNN. The proposed method allows real-time applications using only RGB images while decreasing the hardware requirements. We used the LINEMOD dataset to evaluate the proposed method, and the experimental results show that the proposed method reduces the memory requirement by almost 99% in comparison to the original architecture with the cost of reducing half the accuracy in one of the metrics. Code is available at https://github.com/heitorcfelix/singleshot6Dpose.

This paper presents a method for localizing primitive shapes in a dense point cloud computed by the RGB-D SLAM system. To stably generate a shape map containing only primitive shapes, the primitive shape is incrementally modeled by fusing the shapes estimated at previous frames in the SLAM, so that an accurate shape can be finally generated. Specifically, the history of the fusing process is used to avoid the influence of error accumulation in the SLAM. The point cloud of the shape is then updated by fusing the points in all the previous frames into a single point cloud. In the experimental results, we show that metric primitive modeling in texture-less and unprepared environments can be achieved online.

Finding high-level semantic information from a point cloud is a challenging task, and it can be used in various applications. For instance, it is useful to compactly represent the scene structure and efficiently understand the scene context. This task is even more challenging when using a hand-held monocular visual SLAM system that outputs a noisy sparse point cloud. In order to tackle this issue, we propose an incremental primitive modeling method using both geometric and statistical analyses for such point cloud. The main idea is to select only reliably-modeled shapes by analyzing the geometric relationship between the point cloud and the estimated shapes. Besides that, a statistical evaluation is incorporated to filter wrongly-detected primitives in a noisy point cloud. As a result of this processing, our approach largely improved precision when compared with state of the art methods. We also show the impact of segmenting and representing a scene using primitives instead of a point cloud.

Augmented Reality Authoring Tools are important instruments that can help a widespread use of AR. They can be classified as programming or content design tools in which the latter completely removes the necessity of programming skills to develop an AR solution. Several solutions have been developed in the past years, however there are few works aiming to identify patterns and general models for such tools. This work aims to perform a trend analysis on content design tools in order to identify their functionalities regarding AR, authoring paradigms, deployment strategies and general dataflow models. This work is aimed to assist developers willing to create authoring tools, therefore, it focus on the last three aspects. Thus, 19 tools were analyzed and through this evaluation it were identified two authoring paradigms and two deployment strategies. Moreover, from their combination it was possible to elaborate four generic dataflow models in which every tool could be fit into.

Most work about instructions in Augmented Reality (AR) does not follow established patterns or design rules -- each approach defines its own method on how to convey instructions. This work describes our initial results and experiences towards defining design guidelines for AR instructions. The guidelines were derived from a survey of the most common visualization techniques and instruction types applied in AR. We studied about how 2D and 3D instructions can be applied in the AR context.

According to Gartner's 2013 Hype Cycle for Emerging Technologies, Augmented Reality will reach its Plateu of Productivity before Wearable User Interfaces. In this work, device and user-perspective rendering are compared regarding their applicability to AR-based solutions for Google Glass. The conducted experiment measured and evaluated the advantages and drawbacks on each method and also got positive and negative feedbacks given by users. The tests showed that users preferred the device-perspective approach.

This demonstration allows visitors to use AR applications for monocular optical see-through head-mounted displays with two forms of visualization. One is the device-perspective approach, in which the user sees the virtual content registered with the camera image at the display. The other is the user-perspective method, in which the display is used as a de facto optical see-through device and the virtual content is registered with the real world.

This work presents an improvement to the scene analysis pipeline of a visualization technique called Ghosting. Computer vision and image processing techniques are used to extract natural features, from each video frame. These features will guide the assignment of transparency to pixels, in order to give the ghosting effect, while blending the virtual object into the real scene. Video sequences were obtained from traditional RGB cameras. The main contribution of this work is the inclusion of a motion detection technique to the scene feature analysis step. This procedure leads to a better perception of the augmented scene because the proper ghosting effect is achieved when a moving natural salient object, that catches users attention, passes in front of an augmented one.

This demo presents an AR application that helps the user to solve a jigsaw puzzle that consists of non-textured pieces with a discriminative shape. The pieces are detected, their poses are estimated and the ones that are correctly assembled are highlighted. In order to detect the pieces, the Depth-Assisted Rectification of Contours (DARC) method is used, which performs detection and pose estimation of texture-less planar objects using an RGB-D camera.

In this work in progress we address the problem of interacting with augmented objects. A bare hand tracking technique is developed, which allied to gesture recognition heuristics, enables interaction with augmented objects in an intuitive way. The tracking algorithm uses a flock of features approach that tracks both hands in real time. The interaction occurs by the execution of grasp and release gestures. Physics simulation and photorealistic rendering are added to the pipeline. This way, the tool provides more coherent feedback in order to make the virtual objects look and respond more likely real ones. The pipeline was tested through specific tasks, designed to analyze its performance regarding the easiness of use, precision and response time.

This paper presents a method named Depth-Assisted Rectification of Contours (DARC) for detection and pose estimation of texture-less planar objects using RGB-D cameras. It consists in matching contours extracted from the current image to previously acquired template contours. In order to achieve invariance to rotation, scale and perspective distortions, a rectified representation of the contours is obtained using the available depth information. DARC requires only a single RGB-D image of the planar objects in order to estimate their pose, opposed to some existing approaches that need to capture a number of views of the target object. It also does not require to generate warped versions of the templates, which is commonly needed by existing object detection techniques. It is shown that the DARC method runs in real-time and its detection and pose estimation quality are suitable for augmented reality applications.

This demonstration will allow visitors to use different applications builded for the ARBlocks, a dynamic blocks platform based on projective augmented reality and tangible user interfaces aiming early childhood educational activities. Those applications, along with the platform itself, were designed to be useful tools for educators to teach general subjects for children, such as mathematical and language skills, as well as develop important abilities, like motor coordination and collaboration.

Augmented Reality applications overlap virtual objects over a real scene, taking into account the context, in order to add information to the end user. Nowadays, more advanced applications also make use of Diminished Reality that removes real objects from a scene. This paper describes an approach that combines Augmented Reality and Diminished Reality techniques to modify real objects present in applications. The proposed approach removes an object and replaces it with its purposely-modified replica. The solution uses dynamic texture techniques and Inpaint to enhance the visual response of the modification performed. The results are promising considering both realism of the modified real object and performance of the application.

Augmented Reality applications overlap virtual objects over a real scene, taking into account context, in order to add information to the end user. Nowadays, more advanced applications also make use of Diminished Reality that removes real objects from a scene. This paper describes an approach that combines Augmented Reality and Diminished Reality techniques to modify real objects present in applications. The proposed approach removes an object and replaces it with its purposely-modified replica. The solution uses dynamic texture techniques and Inpaint to enhance the visual response of the modification performed. The results are promising considering both realism of the modified real object and performance of the application.

This paper presents a solution for the photorealistic rendering of synthetic objects into dynamic real scenes, in Augmented Reality applications. In order to achieve this goal, an Image Based Lighting approach is used, where environment maps with different levels of glossiness are generated for each virtual object in the scene at every frame. Due to this, illumination effects, such as color bleeding and specular reflections, can be simulated for virtual objects in a consistent way. A unifying sampling method for the spherical harmonics transformation pass is also used. It is independent of map format and does not need to apply different weights for each sample. The developed technique is combined with an extended version of Lafortune Spatial BRDF, featuring Fresnel effect and an innovative tangent rotation parameterization. The solution is evaluated in various Augmented Reality case studies, where other features like shadowing and lens effects are also exploited.

We introduce a parallel algorithm to solve approximate and exact nearest neighbor queries on the GPU, exploiting its massively parallel processing power. Both data structure construction and nearest neighbor queries are performed on the GPU, avoiding memory copies from system memory to device memory. This algorithm achieves real-time performance, enabling its usage in dynamic scenarios, by minimizing the sorting comparisons needed for a large K value. The underlying data structure for spatial subdivision handles 3D points and is based on grid spatial hashing. Users can specify the grid size interactively. Comparisons were done with other nearest neighbor algorithms implemented on both CPU and GPU. Our approach clearly surpasses CPU implementations regarding processing time, while it presents a competitive solution to GPU ones. Real-time results were obtained with ANN searches (K = 10) for data sets up to 163K points and the potential of our algorithm is demonstrated through a point-based rendering application.

Current GPU computational power enables the execution of complex and parallel algorithms, such as Ray Tracing techniques supported by kD-trees for 3D scene rendering in real time. This work describes in detail the study and implementation of five different kD-Tree traversal algorithms using the parallel framework NVIDIA Compute Unified Device Architecture (CUDA), in order to point their pros and cons regarding adaptation capability to the chosen architecture. In addition, a new algorithm is proposed by the authors based on this analysis, aiming performance improvement. A performance analysis of the implemented techniques demonstrates that two of these algorithms, once adequately adapted to CUDA architecture, are capable of reaching speedup gains up to 15x when compared to former CPU implementations and up to 4x in comparison to existing and optimized parallel ones. As a consequence, interactive frame rates are possible for scenes with 1376x768 pixels of resolution and 1 million primitives.

This paper presents a solution for the photorealistic rendering of synthetic objects into dynamic real scenes, in Augmented Reality applications. In order to achieve this goal, an Image Based Lighting approach is proposed, where environment maps with different levels of glossiness are generated for each virtual object in the scene at every frame. Due to this, illumination effects, such as color bleeding and specular reflections, can be simulated for virtual objects in a consistent way, even under the presence of scene changes. A unifying sampling method for the spherical harmonics transformation pass is also used. It is independent of map format and does not need to apply different weights for each sample. The developed technique is combined with an extended version of Lafortune Spatial BRDF, featuring Fresnel effect and an innovative tangent rotation parameterization. The solution is evaluated in various Augmented Reality case studies, where other features like shadowing and lens effects are also exploited.

This paper presents a markerless tracking technique targeted to the Windows Mobile Pocket PC platform. The primary aim of this work is to allow the development of standalone augmented reality applications for handheld devices based on natural feature tracking of fully 3-Dimensional objects. In order to achieve this goal, a model-based tracking approach that relies on edge information was adopted. Since it does not require high processing power, it is suitable for constrained devices such as handhelds. The OpenGL ES graphics library was used to detect the visible edges in a given frame, taking advantage of graphics hardware acceleration when available. In addition, a subset of two computer vision libraries was ported to the Pocket PC platform in order to provide some required algorithms to the markerless mobile solution. They were also adapted to use fixed-point math, with the purpose of improving the overall performance of the routines. The port of these libraries opens up the possibility of having other computer-vision tasks being executed on mobile platforms. An augmented reality application was created using the implemented technique and evaluations were done regarding tracking performance, accuracy and robustness. In most of the tests, the frame rates obtained are suitable for handheld augmented reality and a reasonable estimation of the object pose was provided.

Single processor technology has been evolving across last decades, but due to physical limitations of chip manufacturing process, the industry is pursuing alternatives to sustain computational power growth,including the creation of multi-core systems. Parallel computing targets problems that are scalable and possibly distributed, dividing the problem into smaller pieces. This approach may be explored to satisfy real-time constraints required by real time physics simulation. This work introduces an optimal solution to implement accurate simulation of deformable bodies in real time. This is accomplished through the use of a meshless simulation technique, known as Point Based Animation, implemented using a parallel framework:the NVIDIA Compute Unified Device Architecture(CUDA). Our implementation showed significative improvements on performance in comparison to the former CPU one: impressive speedups around 20times could be achieved in the simulation of deformable bodies with 575 physical elements (phyxels) and53,504 surface elements (surfels).

Recent years have promoted important evolution in remote sensing, specially taking into consideration quality improvements on satellites that are now capable of obtaining more detailed and accurate observation and producing stereoscopic views. Following market laws, once the level of detail increases the associated costs also rise. In order to take advantage of modest dataset acquisition costs the OpenStereo library (a free, open source, multiplatform, flexible and robust library for stereoscopy) is applied with the purpose of converting satellite images and similar types of geodata acquirers into stereoscopic views.

Digital elevation models may contain several errors, what causes uncertainty about the reliability of the data. Reliable use of elevation data requires that uncertainty associated with the data be accounted for and that the errors responsible for this uncertainty are identified, quantified and removed. Several studies have proposed assorted methods to detect and quantify, and also to remove different kinds of errors. However, these automatic procedures apply algorithms that are specialized in detecting errors with particular characteristics, producing good results only when the model contains predominantly these specific types of errors. In this context, this paper presents a methodology and a tool for enhancing digital elevation models, named DEMEditor. Visual interpretation plays an important role in this work, which exploits user's knowledge about the data in the decision-making process about areas to be enhanced in the digital elevation model. The background of the user allows the identification of any type of error, relieving the need for automatic detection algorithms that specialize in detecting errors with particular characteristics.