Rozdziały

Stały URI dla kolekcjihttp://hdl.handle.net/11652/4775

Przeglądaj

Filtry

20234

Ustawienia

Autorsearch.filters.applied.operator.equals: search.filters.author.Majek, Karol

collection.search.results.head

Teraz wyświetlane 1 - 4 z 4
  • Pozycja
    Using Publicly Available Building Data to Improve 3D Map
    (Wydawnictwo Politechniki Łódzkiej, 2023) Krygiel, Krzysztof; Majek, Karol; Będkowski, Janusz
    In this paper, we address the problem of 3D Map accuracy. No access to RTK GPS or LIDAR leads to poor accuracy of the map. High-rise buildings cause even greater trajectory errors. We used artificial intelligence methods to integrate publicly available building data and show that it can improve map accuracy from monocular camera and inaccurate GPS receiver. The main novelty is a method of building elevation detection in sparse point cloud data. We match detected elevations with building data and use modified bundle-adjustment algorithm to improve the map. We show that proposed approach decreases the trajectory error.
  • Pozycja
    Semantic Segmentation for Autonomous Drone Delivery SUADD’23 Challenge
    (Wydawnictwo Politechniki Łódzkiej, 2023) Mrukwa, Anna; Majek, Karol
    The popularity of drones as well as other different flying devices remains undeterred for several years now, with various industries recognizing their usefulness in a range of applications. However, the effectiveness of such systems is heavily dependent on real-time autonomous surface identification. The goal of this work is to evaluate recently published dataset dedicated to Unmanned Aircraft Systems. We performed experiments using several semantic segmentation neural network architectures. We outline possible improvements for future research and promising results for attentionbased solutions in the field.
  • Pozycja
    BDOT10k-seg: A Dataset for Semantic Segmentation
    (Wydawnictwo Politechniki Łódzkiej, 2023) Kos, Aleksandra; Majek, Karol
    In this work, we describe BDOT10k-seg, a novel aerial dataset for semantic and instance segmentation. Our data covers almost the entire territory of Poland (314,000 km2) and provides precise pixel-level annotations for 286 classes of topographical objects, including buildings, roads, rivers, lakes, airports, agricultural areas, and forests. BDOT10-seg consists of 60,718 images with a resolution of 3 to 75 centimeters per pixel, and more than 40 million object instances. The average image size is 12,367 px because, unlike other publicly available datasets, we do not modify the source orthoimages. The code for generating the BDOT10k-seg dataset is publicly available.
  • Pozycja
    Loss Function Influence on Uncertainty Estimation for White Matter Lesions 3D Segmentation in a Shifted Domain Setting
    (Wydawnictwo Politechniki Łódzkiej, 2023) Kaczmarska, Marta; Majek, Karol
    The aim of this study is to address the problem of distributional shift for white matter Multiple Sclerosis lesion segmentation models. The impact of loss function on models performance and uncertainty estimation is evaluated. The evaluation is performed on two in-domain and one out-ofdomain dataset consisting of 3D FLAIR Magnetic Resonance images. Our experiments show that application of segmentation losses (eg. Dice) translate into reduced models robustness and poorer uncertainty estimation compared with classification losses (eg. CE). The source code is publicly available.