|
|
Differentiable Depth Estimation for Bin Picking
Černý, Marek ; Klusáček, David (advisor) ; Šikudová, Elena (referee)
The goal of this thesis was to investigate the neural 3D surface reconstruction from multiple views with the intent to use the resulting depth maps for bin picking. Survey of papers from 2014 to 2018 showed that none of the state of the art methods would be used to control a robot arm in our setup. Therefore we decided to create our low-level neural approach which we called the EmfNet. The network is based on a pyramidal resolution refining approach. At each pyramid's layer, there are three separate networks that take part in the computation. Each of them has a definite goal, which gives us almost complete understanding of what is going on inside the network. The EmfNet model was partially usable, but we nevertheless extended it to EmfNet-v2. First, another measuring layer was added, which freed EmfNet from depending on an unnecessary hyperparameter. Second, we used constraints on geometry for the network not to be confused by occlusions (cases where a certain part of the surface is visible only from a single camera). Both networks were implemented and tested on a corpus that was created as a part of this thesis. A corpus containing rendered as well as real data. The process of correspondence pairing inside the network can be observed using the visualization tool. We designed a way how to use a robotic arm...
|
|
|
A Calibrated Real-World Colour Picker for Augmented Reality Applications
Zikmund, Martin ; Wilkie, Alexander (advisor) ; Šikudová, Elena (referee)
This thesis describes the process of creating an augmented reality mobile application which allows designers, architects and researchers to retrieve accurate colour information picked from the real-world environment. Specifically, the goal is not only to obtain colour coordinates for the area of interest but to find a set of matching colour samples in an extensive database of colour atlases provided with the application. To properly understand how the camera sensor perceives colour under current lighting conditions, it must be calibrated beforehand using a physical colour chart. Based on this calibration, we can estimate the picked colour coordinates in a standardised colour space. The mobility aspect is the main advantage of the resulting application. Instead of carrying multiple colour atlases, the user can estimate colour matching just using a mobile device and a portable colour chart.
|
|
|
Using neural networks to generate realistic skies
Hojdar, Štěpán ; Křivánek, Jaroslav (advisor) ; Šikudová, Elena (referee)
Environment maps are widely used in several computer graphics fields, such as realistic architectural rendering or computer games as sources of the light in the scene. Obtaining these maps is not easy, since they have to have both a high- dynamic range as well as a high resolution. As a result, they are expensive to make and the supply is limited. Deep neural networks are a widely unexplored research area and have been successfully used for generating complex and realistic images like human portraits. Neural networks perform well at predicting data from complex models, which are easily observable, such as photos of the real world. This thesis explores the idea of generating physically plausible environment maps by utilizing deep neural networks known as generative adversarial networks. Since a skydome dataset is not publicly available, we develop a scalable capture process with both low-end and high-end hardware. We implement a pipeline to process the captured data before feeding it to a network and extend an already existing network architecture to generate HDR environment maps. We then run a series of experiments to determine the quality of the results and uncover the directions of possible further research.
|
|
|
3D object classification using neural networks
Krabec, Miroslav ; Křivánek, Jaroslav (advisor) ; Šikudová, Elena (referee)
3D Object Classification Using Neural Networks Bc. Miroslav Krabec Classification of 3D objects is of great interest in the field of artificial intelligence. There are numerous approaches using artificial neural networks to address this problem. They differ mainly in the representation of the 3D model used as input and the network architecture. The goal of this thesis is to explore and test these approaches on publicly available datasets and subject them to independent comparison, which has not so far appeared in the literature. We provide a unified framework allowing to convert the data from common 3D formats. We train and test ten different network on the ModelNet40 and ShapeNetCore datasets. All the networks performed reasonably well in our tests, but we were generally unable to achieve the accuracies reported in the original papers. We suspect this could be due to extensive, albeit unreported, hyperparameter tuning by the authors of the original papers, suggesting this issue would benefit from further research. 1
|
|
|
Optical analysis of pellet car damages
Dubský, Jan ; Šikudová, Elena (advisor) ; Škovierová, Júlia (referee)
During iron processing, pelletizing is one of the necessary steps. Iron ore pellets are burned on pellet burning line, which consists of individual pellet cars. Due to constant temperature changes, pellet cars get damaged. This thesis focuses on an optical analysis of pellet car damages. The al- gorithm presented can find individual ribs of pellet car and perform analysis of their shape, position and look. Such analysis can provide basic information for damage evaluation, same as the material for further research of pellet car damage causes. 1
|
|
|
Klasifikace na množinách bodů v 3D
Střelský, Jakub ; Mráz, František (advisor) ; Šikudová, Elena (referee)
Increasing interest for classification of 3D geometrical data has led to discov- ery of PointNet, which is a neural network architecture capable of processing un- ordered point sets. This thesis explores several methods of utilizing conventional point features within PointNet and their impact on classification. Classification performance of the presented methods was experimentally evaluated and com- pared with a baseline PointNet model on four different datasets. The results of the experiments suggest that some of the considered features can improve clas- sification effectiveness of PointNet on difficult datasets with objects that are not aligned into canonical orientation. In particular, the well known spin image rep- resentations can be employed successfully and reliably within PointNet. Further- more, a feature-based alternative to spatial transformer, which is a sub-network of PointNet responsible for aligning misaligned objects into canonical orientation, have been introduced. Additional experiments demonstrate that the alternative might be competitive with spatial transformer on challenging datasets. 1
|
|
|
Evaluation of Dynamic Range Reconstruction Approaches and a Mobile Application for HDR Photo Capture
Mirbauer, Martin ; Křivánek, Jaroslav (advisor) ; Šikudová, Elena (referee)
Digital photography became widespread with the global use of smartphones. However, most of the captured images do not fully use the camera capabilities by storing the captured photos in a format with limited dynamic range. The subject of dynamic range expansion and reconstruction has been researched since early 2000s and recently gave rise to several new reconstruction methods using convolutional neural networks (CNNs), whose performance has not yet been comprehensively compared. By implementing and using our dynamic range reconstruction evaluation framework we compare the reconstruction quality of individual CNN-based approaches. We also implement a mobile HDR camera application and evaluate the feasibility of running the best-performing reconstruction method directly on a mobile device.
|
|
|
Automated number plate recognition from low quality video-sequences
Vašek, Vojtěch ; Franc, Vojtěch (advisor) ; Šikudová, Elena (referee)
The commercially used automated number plate recognition (ANPR) sys- tems constitute a mature technology which relies on dedicated industrial cam- eras capable of capturing high-quality still images. In contrast, the problem of ANPR from low-quality video sequences has been so far severely under- explored. This thesis proposes a trainable convolutional neural network (CNN) with a novel architecture which can efficiently recognize number plates from low-quality videos of arbitrary length. The proposed network is experimentally shown to outperform several existing approaches dealing with video-sequences, state-of-the-art commercial ANPR system as well as the human ability to recog- nize number plates from low-resolution images. The second contribution of the thesis is a semi-automatic pipeline which was used to create a novel database containing annotated sequences of challenging low-resolution number plate im- ages. The third contribution is a novel CNN based generator of super-resolution number plate images. The generator translates the input low-resolution image into its high-quality counterpart which preserves the structure of the input and depicts the same string which was previously predicted from a video-sequence. 1
|
|
| |
|
|
Local sharpness prediction and image segmentation
Kopál, Jakub ; Šikudová, Elena (advisor) ; Horáček, Jan (referee)
The problem of automatic segmentation turned out to be complicated andtothisday, notcompletelysolved.Sinceitisacomplexproblem,thispaperis not- tryingtosolveitinitsmostgeneralform. Instead,itisfocusedonautomatic, bina- rypicturesegmentation,withtheoptiontochooseattributes, basedonwhich the seg- mentation should operate. Among these attributes are the focus and color of the picture. The results of the segmentation based on the assumption "focused object, blurry background" turned out to be very similar to the groundtruth in pictures, which fulfill this assumption. 1
|
guest ::
