Search

Image Inpainting, the task of recovering missing information in an image, is of great practical interest to users who wish to recover damaged photographs or remove unwanted parts of images. Advances in Convolutional Neural Networks and Adversarial Networks have produced large gains in this problem. Nevertheless, Inpainting algorithms may fail to recover a decent image, and even produce jarring artifacts. In this work, we seek to uncover what causes such issues. We investigate the relationship between the artifacts and the type of degradation mask, the size of the mask, and the algorithms used. To that end, we employ rectangular, free-form and segmentation-based degradation masks, of varying sizes. We find that large, continuous regions of missing information are often to blame for the most egregious failures, and not really how complex the shape of the region is. Moreover, we observe that no algorithm is completely immune to these failure modes.

3D reconstruction is one of the main topics in computer vision and is heavily applied for creating virtual environments. Photogrammetry is a technique for obtaining 3D information by mapping objects and scenarios using only images. However, this process can take a long time when using large datasets. In this paper, a permissive open-source pipeline is proposed focusing on robustness, efficiency, and low execution time in batch processing. The permissive license allows commercial use without the need of keeping the code open. We mixed an enhanced structure from motion algorithm and a recurrent multi-view reconstruction. We also use Point Cloud Library for normal estimation, surface reconstruction, and texture mapping. We compared our results with COLMAP and MVE techniques using the DTU MVS dataset and real-world scenarios with our own datasets. The results showed a decrease of 69.4% on average time (compared to the best result of other techniques), but also demonstrated the need for more images to generate a complete reconstructed model.

Image-based body tracking algorithms are useful in several scenarios, such as avatar animations and gesture interaction for VR applications. In the last few years, the best-ranked solutions presented on the state of the art of body tracking (according to the most popular datasets in the field) are intensively based on Convolutional Neural Networks (CNNs) algorithms and use large datasets for training and validation. Although these solutions achieve high precision scores while evaluated with some of these datasets, there are particular tracking challenges (for example, upside-down cases) that are not well-modeled and, therefore, not correctly tracked. Instead of lurking an all-in-one solution for all cases, we propose HuTrain, a framework for creating datasets quickly and easily. HuTrain comprises a series of steps, including automatic camera calibration, refined human pose estimation, and known dataset formats conversion. We show that, with our system, the user can generate human pose datasets, targeting specific tracking challenges for the desired application context, with no need to annotate human pose instances manually.

Gestural interfaces in virtual reality (VR) expand the design space for user interaction, allowing spatial metaphors with the environment and more natural and immersive experiences. Typically, machine learning approaches recognize gestures with models that rely on a large number of samples for the training phase, which is an obstacle for rapidly prototyping gestural interactions. In this paper, we propose a solution designed for hi-fi prototyping of gestures within a virtual reality environment through a high-level Domain-Specific Language (DSL), as a subset of the natural language. The proposed DSL allows non-programmer users to intuitively describe a broad domain of poses and connect them for compound gestures. Our DSL was designed to be general enough for multiple input classes, such as body tracking, hand tracking, head movement, motion controllers, and buttons. We tested our solution for wands with VR designers and developers. Results showed that the tool gives non-programmers the ability to prototype gestures with ease and refine its recognition within a few minutes.

Through the last decade, the creation of extended reality (XR) solutions has significantly increased due to the advent of cheaper, more advanced, and accessible instruments like smartphones, headsets, platforms, development kits, and engines. For instance, the number of GitHub repositories for XR related projects jumped from 51 in 2010 to over 15,000 in 2020. At the same time, the developer community approaches the creation of XR applications using inherited design processes and methods from past mainstream platforms such as web, mobile, or even product design. Unfortunately, those platforms do not consider the spatial aspects of these applications. In this paper, we present a revisited design process and a toolkit focused on the challenges innate to XR, that aims to help beginners and experienced teams in the creation of applications and interactions in Virtual, Augmented, and Mixed Reality. We also present a compendium of 113 techniques and 118 guidelines and a set of canvases that guides users through the process, preventing them from skipping important tasks and discoveries. At last, we present a pilot case where we accompany a team with developers and designers running our process and using our toolkit for the first time, showing the benefits of a process that strikes specific issues of XR apps.

Until a few years ago, internationalization was seen as a self-feeding process, resulting from decisions based on accumulated knowledge. De-internationalization is not an option. However, the reality shows us that enterprises have done international divestment. To this date, the research on de-internationalization remains somehow scarce and needs additional research. The authors want to understand the causes of de-internationalization. They use the grounded theory methodology, studying six Portuguese cases that already have a process of de-internationalization. Five firms cooperate with, presenting six cases. The authors distinguish causes that occur before/after the internationalization process begin and causes related to firm or external to firm. One of the main causes of de-internationalization is the incorrect internationalization process. The Misfit Micro-Location (the wrong choice of place inside a determined country) is a very important cause, which means that, after choosing the country, it is necessary to choose the best place to be located.

In the past years, captures made by dual-fisheye lens cameras have been used for virtual reality, 360 broadcasting and many other applications. For these scenarios, to provide a good- quality experience, the alignment of the boundaries between the two images to be stitched must be done properly. However, due to the peculiar design of dual-fisheye cameras and the high variance between different captured scenes, the stitching process can be very challenging. In this work, we present a 360 stitching solution based on feature cluster matching. It is an adaptive stitching technique based on the extraction of feature cluster templates from the stitching region. It is proposed an alignment based on the template matching of these clusters, successfully reducing the discontinuities in the full-view panorama. We evaluate our method on a dataset built from captures made with an existing camera of this kind, the Samsung's Gear 360. It is also described how we can extend these concepts from image stitching to video stitching using the temporal information of the media. Finally, we show that our matching method outperforms a state-of-the-art matching technique for image and video stitching.

In this work we address the problem of landmark recognition. We extend PlaNet, a model based on deep neural networks that approaches the problem of landmark recognition as a classification problem and performs the recognition of places around the world. We propose an extension of the PlaNet technique in which we use a voting scheme to perform the classification, dividing the image into previously defined regions and inferring the landmark based on these regions. The prediction of the model depends not only on the information of the features learned by the deep convolutional neural network architecture during training, but also uses local information from each region in the image for which the classification is made. To validate our proposal, we performed the training of the original PlaNet model and our variation using a database built with images from Flickr, and evaluated the models in the Paris and Oxford Buildings datasets. It was possible to notice that the addition of image division and voting structure improves the accuracy result of the model by 5-11 percentage points on average, reducing the level of ambiguity found during the inference of the model.

Due to the increase in smartphone usage, it became easier to register memorable moments with a more accessible camera. To ensure a nice capture was made, users often take multiple shots from a scene, later filtering them based on some quality criteria. However, sometimes this may be unfeasible to do manually. To address this issue, this work initially defines relevant characteristics present in a good personal picture or video. We then show how to automatically search for these aspects using computer vision algorithms, successfully assessing personal media based on these aspects. Moreover, we show that it was possible to use this proposed solution in a real-world application, improving the generation of a personal album containing the best pictures and videos.