|
|
Fusing image and non-grid-like data for object segmentation
Repka, Samuel ; Nosko, Svetozár (oponent) ; Zemčík, Pavel (vedoucí práce)
Quite often, a phenomenon of interest can described by more than one data source. For example, a car's appearance shows its colour and brand, but not its engine status. However, other data sources do provide us with this information, be it a sound or mere touch. Such data source is often referred to as a modality. While using a single data source to extract the needed information may be sufficient, the addition of more modalities can be beneficial, because of their complementary nature. This data fusion, however, may be a quite challenging process. Different kinds of data have different properties, structures and various challenges connected to them. A plethora of different methods has been proposed, but usually, the methods are very data-dependent. This thesis presents a new approach to the fusion of two modalities, primarily for the purpose of image segmentation. One of the modalities is image, and the second one is any non-grid-like modality. The method uses a graph to jointly represent both modalities, aiming to capture the intra and inter-modalities relationships as accurately as possible. The graph is then processed, producing a graph with fused data, or a direct segmentation. The proposed method was evaluated on two datasets (from the fields of mineralogy and timber processing) and compared to another solution, showing both the potential and limitations of the method. In case of the mineralogy dataset, the results are very encouraging, showing that the method is capable of data fusion, even outperforming a contemporary method. In case of the timber dataset, the results were not as conclusive, as the method failed to improve the results when compared to a baseline solution, which may have been caused by a challenging dataset.
|
|
|
Machine Learning-Based Multimodal Data Processing and Mapping in Robotics
Ligocki, Adam ; Duchoň,, František (oponent) ; Saska,, Martin (oponent) ; Žalud, Luděk (vedoucí práce)
This dissertation deals with the application of object detection neural networks on multimodal data in robotics. It aims at three topics: dataset-making, multimodal data processing, and neural network training. The most important is a proposed method that allows creating a large training dataset without an expensive and time-demanding human annotation. The method uses the neural network model trained on the RGB image data and uses multiple sensors' data to create the surrounding map and transfers the annotations of objects detected in the RGB image to the other data domain, like thermal images or point cloud data. Applying this approach, the author generated the thermal image dataset, which contained hundreds of thousands of annotated images, and used them to train the network that outperformed other models trained on human-annotated data. Moreover, the thesis also studies the robustness of object detection in various data domains during difficult weather conditions. The thesis also describes the entire multimodal data processing pipeline that the author created during his Ph.D. studies. That includes developing a unique sensory framework that employs a wide range of commonly used sensors in robotics and self-driving cars. Next, it describes the process of using the sensory framework to make a large-scale publically available open-source navigation and mapping dataset called Brno Urban Dataset. Finally, it covers the description of the custom-made software tools, the Atlas Fusion and the Robotic Template Libarary that the author used to manipulate the multimodal data.
|
|
|
Machine Learning-Based Multimodal Data Processing and Mapping in Robotics
Ligocki, Adam ; Duchoň,, František (oponent) ; Saska,, Martin (oponent) ; Žalud, Luděk (vedoucí práce)
This dissertation deals with the application of object detection neural networks on multimodal data in robotics. It aims at three topics: dataset-making, multimodal data processing, and neural network training. The most important is a proposed method that allows creating a large training dataset without an expensive and time-demanding human annotation. The method uses the neural network model trained on the RGB image data and uses multiple sensors' data to create the surrounding map and transfers the annotations of objects detected in the RGB image to the other data domain, like thermal images or point cloud data. Applying this approach, the author generated the thermal image dataset, which contained hundreds of thousands of annotated images, and used them to train the network that outperformed other models trained on human-annotated data. Moreover, the thesis also studies the robustness of object detection in various data domains during difficult weather conditions. The thesis also describes the entire multimodal data processing pipeline that the author created during his Ph.D. studies. That includes developing a unique sensory framework that employs a wide range of commonly used sensors in robotics and self-driving cars. Next, it describes the process of using the sensory framework to make a large-scale publically available open-source navigation and mapping dataset called Brno Urban Dataset. Finally, it covers the description of the custom-made software tools, the Atlas Fusion and the Robotic Template Libarary that the author used to manipulate the multimodal data.
|
|
|
Metody lokalizace rozdílů v různých modálitách malířských děl
Fürbach, Radek ; Blažek, Jan (vedoucí práce) ; Křivánek, Jaroslav (oponent)
Práce se zabývá analýzou malířských děl za účelem zjištění použitých malířských technik. Konkrétně se zaměřuje na lokalizaci podkladových kreseb na základě porovnání snímků pořízených ve spektrech s rozdílnou penetrační hloubkou. Definuje problém spojený se zachycením porovnávaných snímků v různých spektrech. Specifikuje metody, které určují závislost mezi částmi spektra (převážně RGB a IR) a na základě zjištěné závislosti aproximují převod mezi těmito částmi spektra (Projekce červené složky spektra, Intenzita barvy, Vážený průměr složek spektra, Tabulkový přepočet, Lineární regrese, PCA analýza a Hranová dekompozice). Práce též popisuje obecnější problémy znesnadňující řešení dané úlohy, jako je šum, nerovnoměrné osvětlení a sčítání stejného typu záření. Problémy jsou v práci důkladně rozebrány. Navrhujeme Výpočet parametrů osvětlení pomocí neuronové sítě, Aproximace intenzity osvětlení rozmazáním, Aproximace intenzity osvětlení polynomem a Aproximace intenzity osvětlení metodou TWMJ pro potlačení nerovnoměrného osvětlení. Definujeme metody Odhadnutí z hranové dekompozice a Lokální metoda nejmenších čtverců řešící sčítání stejného typu záření. Dále popisujeme Gaussův filtr, Průměrování, Mediánový filtr, Konzervativní vyhlazení a Průměrování s mezí pro odstranění šumu. Navržené metody jsou experimentálně...
|
host ::
RSS.