Browsing by Author "Kannala, Juho"

Filter results by typing the first few letters
Now showing 1 - 20 of 218
  • 3D Gaussian splatting theory and variance rendering extension
    (2024-09-12) Taka, Veikka
     Perustieteiden korkeakoulu | Master's thesis
    Radiance field methods are a recent approach for novel-view synthesis. Radiance field methods reconstruct a representation of a scene from a set of images that capture the scene from multiple viewpoints. This scene representation can then be used to render photorealistic images from novel viewpoints. In 3D Gaussian Splatting (3DGS), the scene representation consists of hundreds of thousands of overlapping multivariate Gaussian distributions in three-dimensional space, centred on the discrete point samples of the underlying continuous density distribution. An image is rendered by projecting these Gaussians into the camera's image plane, sampling the 2D Gaussian projections at each pixel location, and compositing the sample values in depth order. 3DGS addresses the limitations of the preceding state-of-the-art method, namely Neural Radiance Fields, facilitating rapid reconstruction of the scene representation and real-time rendering. This thesis introduces a statistical formulation of the splatting equation, the image formation model of 3DGS, as a weighted arithmetic mean. This allows the compositing of generic point cloud quantities, such as depth, into a properly normalised image. Furthermore, it permits the rendering of a variance image, which describes the variance of the rendered Gaussian quantity at each pixel. The mathematical formulae and pseudocode are provided for the implementation of the variance image renderer and for the computation of analytical gradients through the variance image back to the scene parameters. Furthermore, a variance image renderer has been implemented on top of Splatfacto, an open-source implementation of 3DGS.
  • 5G ja sen vastaaminen IoT:n tarpeisiin
    (2016-12-17) Toivonen, Jussi
     Perustieteiden korkeakoulu | Bachelor's thesis
  • 5G-verkkoteknologian kehitys ja vaikutus esineiden internetiin
    (2018-04-25) Helkavaara, Ilkka
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Adaptiivinen videon suoratoisto
    (2016-12-18) Vartiainen, Aleksi
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Adversiaalinen haku täydellisen tiedon peleissä
    (2023-12-15) Pollari, Visa
     Perustieteiden korkeakoulu | Bachelor's thesis
  • AI for Platform Economy
    (2018-04-25) Sipilä, Pekka
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Akustisten tapahtumien reaaliaikainen tunnistus
    (2017-12-22) Rouvinen, Miika
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Algorithmic design of RNA nanostructures
    (2019-04-24) Virtanen, Raili
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Algoritmeja likimääräiseen bayesilaiseen laskentaan
    (2017-01-04) Kähkönen, Erik
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Algoritmien visualisointijärjestelmät ja niiden vaikutukset oppimiseen
    (2017-09-16) Mustonen, Miika
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Angular 2
    (2017-04-16) Hassanzadeh, Pejam
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Apache Spark 2.0 tehokkuuden näkökulmasta
    (2016-12-18) Arasalo, Ossi
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Application of Computational Histopathology to Quantify Immune Cells in Clear-cell Renal Cell Carcinoma Tissue
    (2022-01-17) Brummer, Otso
     Perustieteiden korkeakoulu | Master's thesis
    Spatial arrangement and number of lymphocytes in biopsy textures have been shown to influence the prognosis and the treatment efficiency in cancers. In particular, the intratumoral lymphocyte infiltration generally correlates positively with prognosis. In the case of renal cell clear cell carcinoma however, the association is ambiguous. Previously infeasible clinical problems, such as the lymphocyte infiltration quantification, may be computationally feasible with the application of modern convolutional neural networks. This thesis aims to produce a combination of two quantification networks which can recognize the texture and number of lymphocytes in whole slide renal cell clear cell carcinoma nephrectomy samples. Moreover, a preliminary attempt at induction of the immune response phenotypes from the results is made. An annotated dataset consisting of 36 000 labelled texture images and a regional dataset of 200 images with labelled lymphocytes and general cells were produced from the publicly available renal cell carcinoma hematoxylin-eosin-stained biopsy scans of Cancer Genome Atlas image archive. These datasets were utilized to train two convolutional networks. One to evaluate the tissue texture and one to evaluate the type and counts of cells in an image. The tissue texture evaluation network achieved total accuracy of 90 % in a collected testing dataset. Often the tissue is a combination of multiple textures and defining a single label for it without a degree of ambiguity is impossible. Taking this into account, the results are more than satisfactory. The cell classification network achieved VOC AUC 0.5 of 62.3 % for non-lymphocytes and 48.7% for lymphocytes. Cell classification suffers considerable from the variance between different images. Localization of the cell matches is satisfactory, but the differentiation of lymphocytes from other cells mainly depends on the size and darkness of the cell in a stain which has an amount of volatility across different staining protocols. Otherwise, the results are comparable to previous studies conducted on the immune cell counts. This thesis indicates that computational analysis of histopathological images is possible even with consumer grade systems. A set of stain color normalization solutions could be developed to reliably produce uniform images to ease the challenging lymphocyte classification problem.
  • Applying Large-Scale Image Retrieval to Near-Duplicate Image Detection
    (2016-10-27) Ulmanen, Samuli
     Sähkötekniikan korkeakoulu | Master's thesis
    Perceptual hashing outputs an image identifier that can be used for detecting images similar to the original image also known as near-duplicate images. ThingLink is a commercial image annotation service using perceptual hashing for placing annotations from the original annotated image to near-duplicate images. A customer is reporting 20-30% of near-duplicate images missing annotations. The system is working as expected calling for improved near-duplicate image detection (NDID) methodology. We apply Local Features and large-scale image retrieval to near- duplicate image detection. We use a Bag-of-Visual-Words-based image retrieval system for near-duplicate detection by assuming the original image always has the highest score of the images returned by Bag-Of-Visual-Words query. The query always returns the best matching image regardless of how good the match. We employ a cutoff score and classify all queries returning images with scores below the cutoff as no duplicate found. We show the Local Features and large- scale image retrieval system is better than the perceptual hash-based systems by generating seven different types of near-duplicate image sets from original images in two datasets. The originals form the image database. In addition we use a set of predicted images not in the database to determine how well the systems classify queries as no duplicate found. We show the optimal cutoff score to be the maximum score returned while querying predicted negative images for a given dataset. For matching near-duplicates the perceptual hashing schemes use the Hamming Distance, the number of bits by which hashes differ. We find an optimal Hamming Distances for both hashes. Despite tuning, we demonstrate Local Features and large-scale image retrieval to be the superior system for both datasets and all seven types of near-duplicate images used in near-duplicate image detection simulations.
  • Artificial intelligence technologies in automated vehicles
    (2019-04-21) Alkio, Kyösti
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Automated structure discovery in atomic force microscopy
    (2020-02-26) Alldritt, Benjamin; Hapala, Hapala; Oinonen, Niko; Urtev, Fedor; Krejci, Ondrej; Federici Canova, Filippo; Kannala, Juho; Schulz, Fabian; Liljeroth, Peter; Foster, Adam
     A1 Alkuperäisartikkeli tieteellisessä aikakauslehdessä
    Atomic force microscopy (AFM) with molecule-functionalized tips has emerged as the primary experimental tech- nique for probing the atomic structure of organic molecules on surfaces. Most experiments have been limited to nearly planar aromatic molecules due to difficulties with interpretation of highly distorted AFM images originat- ing from nonplanar molecules. Here, we develop a deep learning infrastructure that matches a set of AFM images with a unique descriptor characterizing the molecular configuration, allowing us to predict the molecular struc- ture directly. We apply this methodology to resolve several distinct adsorption configurations of 1S-camphor on Cu(111) based on low-temperature AFM measurements. This approach will open the door to applying high-resolution AFM to a large variety of systems, for which routine atomic and chemical structural resolution on the level of individual objects/molecules would be a major breakthrough.
  • Automated Tip Functionalization via Machine Learning in Scanning Probe Microscopy
    (2022-04) Alldritt, Benjamin; Urtev, Fedor; Oinonen, Niko; Aapro, Markus; Kannala, Juho; Liljeroth, Peter; Foster, Adam S.
     A1 Alkuperäisartikkeli tieteellisessä aikakauslehdessä
    Auto-CO-AFM is an open-source software package for scanning probe microscopes that enables the automatic functionalization of scanning probe tips with carbon monoxide molecules. This enables machine operators to specify the quality of the tip needed utilizing a pre-trained library with off-the-shelf software. From a single image, the software package can determine which molecules on a surface are carbon monoxide, perform the necessary tip functionalization procedures, interface with microscope software to control the tip position, and determine the centeredness of the tip after a successful functionalization. This is of particular interest for atomic force microscopy imaging of molecules on surfaces, where the tip functionalization is a necessary and time consuming step needed for sub-molecular resolution imaging. This package is freely available under the MIT License.
  • Automatic Floorplan Analysis
    (2019-12-16) Kalervo, Ahti
     Perustieteiden korkeakoulu | Master's thesis
    Digital representation of a housing floorplan is a must in a property advertisment these days. The property advertisment might contain either a two dimensional floorplan or then a three dimensional interior model of the apartment. Modern computer vision methods are able to reproduce a 2D or a 3D model from a scanned floorplan using semantic segmentation and then redraw it using a vector graphic representation. A vectorized floorplan can be easily edited further if the scanned floorplan contained some mistakes. Modern convolutional neural network models have been show to work well in many image recognition task such as object detection and semantic segmentation. However, neural networks have not been widely used in floorplan analysis. Previous methods have been using approaches based on strong heuristics and low-level pixel operations. One reason for this is that there is a clear lack of large enough dataset, which is a prerequisite for deep neural networks. This thesis presents an image dataset called CubiCasa5K and a method for floorplan parsing that relies on a multi-task convolutional neural network. The floorplan image dataset contains 5000 samples annotated into over 80 floorplan object categories. The dataset annotations are performed in a dense and versatile manner by using polygons for separating the different objects. The dataset is five times bigger that the previous biggest data set and the quality of the annotations is higher than reported previously. The new deep learning model requires less post-processing after the neural network has done the prediction. Also the model able to reproduce and improve the previous state of the art results in automatic floorplan analysis. Furthermore, we apply the same state of the art model to the new CubiCasa5K dataset and present the first results from the new dataset. After this we still improve our results by changing the models internal data representation so that the mode is able to predict diagonal walls. By releasing the novel dataset and parts of our implementations, this study significantly boosts the research on automatic floorplan image analysis as it provides a richer set of tools for investigating the problem in a more comprehensive manner.
  • Automatic object detection from images
    (2017-02-04) Eskonen, Juha
     Perustieteiden korkeakoulu | Bachelor's thesis
  • Automatic segmentation workflow of dark features for oil spill monitoring using synthetic aperture radar
    (2021-01-25) Llop Cardenal, Roberto
     Perustieteiden korkeakoulu | Master's thesis