|
|
Material picker: Material recognition in images using machine learning
Jurčák, Filip ; Vévoda, Petr (advisor) ; Šikudová, Elena (referee)
The process of setting material properties for realistic appearance after rendering is usually tiresome and often requires carefully crafted skill for fine-tuning the parameters, as different combinations of these parameters can produce different-looking materials. To simplify this process, we introduce a solution to the texture transfer problem by creating a pipeline containing several deep neural networks. These networks subsequently represent solutions to inverse rendering and material segmentation by predicting intrinsic scene characteristics, like diffuse and specular albedo, surface normals, glossiness, view vector, texture coordinates, and segmentation, all from a single image. Artists can subsequently plug these inferenced properties inside their 3D scene representations and thus reduce the time needed to iterate over several design ideas. To train these networks, we generated high-quality dataset of substantial size using physically-based techniques to ensure good generalization on real-world images. 1
|
|
|
6D pose estimation of objects in images
Cífka, Martin ; Šivic, Josef (advisor) ; Šikudová, Elena (referee)
The 6D pose estimation is an important computer vision task with applications in robotics, e.g. for manipulation or grasping, but also in computer graphics and augmented reality. Given an image, the task is to estimate the 3D rotation and 3D translation of the known object with respect to the camera. The task is even more challenging in an uncontrolled environment, e.g. when we do not have proper camera calibration. In that case, the focal length also needs to be estimated with the 6D pose. In this work, we address the issues of methods that work in such uncontrolled environments. First, we focus on FocalPose, a state-of-the-art method for joint estimation of object 6D pose and camera focal length. We review the method and propose several improve- ments. These include (i) re-deriving and improving the 6D pose and focal length update rule, (ii) replacing the model retrieval method, and (iii) changing the distribution of 6D poses and focal lengths used for synthetic training data rendering. These changes lead to improved results compared to the state-of-the-art FocalPose method. Second, to avoid often costly retraining of models for 6D pose estaimation, it is ben- eficial to consider methods with the ability to generalize to novel objects that have not been seen during training. These methods require a 2D...
|
|
|
Weather Webcam Calibration
Krsička, Ondřej ; Mirbauer, Martin (advisor) ; Šikudová, Elena (referee)
The work deals with the calibration of meteorological web cameras, which can be then used for training of neural networks for the creation of realistic lighting and the genera- tion of realistic clouds in the sky into 3D scenes. A method of calibration using the sun position, a method using the sky appearance and a simple analytical sky model and a method based on neural networks are compared. The first two methods are implemented in Python. A more sophisticated sky model is incorporated into the sky appearance cali- bration method and compared to the original sky model and the sun position calibration. Sun position calibration is the most successful, calibration with the original sky model a little less. The more sophisticated sky model doesn't improve the calibration the way it is used and further improvements are suggested. 1
|
|
| |
|
|
Generative neural networks for sky image outpainting
Mrázek, Matěj ; Šikudová, Elena (advisor) ; Mirbauer, Martin (referee)
Image outpainting is a task in the area of generative artificial intelligence, where the goal is to expand an image in a feasible way. The goal of this work is to create a machine learning algorithm capable of sky image outpainting by implementing sev- eral recently proposed techniques in the field. We train three models, a tokenizer for converting images to tokens and back, a masked generative transformer for performing outpainting on tokens and a super sampler for upscaling the result, all on a dataset of sky images. Then, we propose a procedure that combines the trained models to solve the outpainting task. We describe the results of training each model and those of the fi- nal algorithm. Our contribution consists mainly in providing a working, open-source implementation including the trained models capable of sky image outpainting. 1
|
|
|
Creating 3D Diorama from Single Image with Deep Learning
Vejbora, Martin ; Šikudová, Elena (advisor) ; Holeňa, Martin (referee)
The goal of this thesis is to automate the process of generating 3D dio- rama scenes from a single image. After an extensive analysis of existing approaches, we propose to combine the output of deep learning models for panoptic segmentation and monocular depth estimation. We encountered some limitations of the available depth model for our use case, which we addressed through fine-tuning. To construct the diorama, we separate the objects identified by segmentation into distinct images with transparent back- grounds. These images are placed in a 3D scene, arranged in a way that reflects the estimated depth of each object. We implemented our method as an add-on for Blender. The thesis was developed in collaboration with a company called polygoniq.
|
|
|
Decoding visual stimuli from cortical activity
Vašek, Vojtěch ; Antolík, Ján (advisor) ; Šikudová, Elena (referee)
This thesis aims to develop a machine learning model that can decode stimu- lus images from cortical activity in the primary visual cortex (V1) to understand the relationship between V1 activity and visual perception. The limited avai- lability of biological data makes it necessary to use the spiking neural network model of V1 to generate the underlying training data. Machine learning tech- niques, particularly neural networks, will be explored to generate high-quality stimulus images. Standard loss functions, as well as discriminator loss from GAN networks training, will be used to train the decoding models. Linear regression models will be used baseline. The research questions to be addressed include the best decoding approach, the impact of the number of neurons recorded or stimuli presented, the loss of information in high frequencies domain and the effect of intrinsic noise in neural responses on reconstructing visual stimuli. This thesis proposes a trainable convolutional network, which outperforms other baseline models such as linear regression. We observe that the loss function producing the best results is the MSSSIM. However, the intrinsic noise in neural respon- ses limits the reconstruction, and only low frequencies are being reconstructed. The size of the dataset and the number of cortical...
|
|
|
Automatic analysis of squash straight drives accuracy from a single camera view
Veedla, Walter Herold ; Šikudová, Elena (advisor) ; Goliaš, Matúš (referee)
Squash is a racket and ball sport with an estimated 20 million players worldwide. Compared to sports like tennis and golf, squash tracking and analysis systems are rela- tively underdeveloped and performance analysis is often done by manual instruction or by pencil-and-paper. While in the recent years more advanced squash specific technology has become available, it requires high-cost specialised hardware and does not capture the location of the bounce of the ball on the floor. This project attempts to tackle this gap of existing squash analysis tools by using computer vision techniques to automate the collection of shot data of a common squash training drill "straight drives", where the ball is being repeatedly hit parallel to a side-wall of the court. An analytics program is developed that can process a video file of a player performing the "straight drives" drill and produce accuracy metrics from the video. The result of this work is a computer program that allows an easy way for the user to get feedback from their training and track their progress. 1
|
|
|
Application for modeling buildings from photographs
Medek, Jakub ; Ježek, Pavel (advisor) ; Šikudová, Elena (referee)
Title: Application for modeling buildings from photographs Author: Jakub Medek Department: Department of Distributed and Dependable Systems Supervisor: Mgr. Pavel Ježek, Ph.D., Department of Distributed and Dependable Systems Abstract: The goal of the thesis was to develop a Windows application that can be used for modelling objects based on their photographs. The user can import photos of an object into the application and then start creating the model in two steps. First, the user has to calibrate a virtual camera for each photo. After that, the user can create the 3D model by drawing edges and faces on top of the photo. The application supports exporting the created model in a standard 3D format .obj including textures generated from photos. The application is written in .NET Core framework with the WPF graphical framework. WPF and one of the used libraries Avalondock are located on the View layer. The rest of the application (ViewModel and Model layers) use only cross- platform libraries. This simplifies a possibility of future change of the graphical framework to make the application cross-platform. Keywords: .NET Core WPF 3D modelling Photo matching Texture projection
|
|
|
Panorama stitching regularly spaced photos
Douša, Marek ; Šikudová, Elena (advisor) ; Goliaš, Matúš (referee)
This thesis addresses the problem of stitching photos into a panorama in two ways - using features and uniformly distributed photos. The mathematical basis is introduced in the form of transformations used in both approaches, followed by the currently mainstream approach to stitching, and finally the theoretical part describes the evenly spaced image approach, for which a demonstrative program is also provided. Both approaches have their advantages and disadvantages on which the thesis emphasizes. This bachelor thesis is designed to serve as a stepping stone for anyone who would like to study the topic of stitching photos into panoramas along with the current state of the art, which has its pros and cons, so that the reader can avoid dead ends or possible surprises during the creation of panoramas.
|
guest ::
