Raimund Schnürer

Last Name

Schnürer

First Name

Raimund

ORCID

0000-0002-7558-3621

Show all metadata

Search Results

Publications 1 - 10 of 25

The Next Generation of Atlas User Interfaces
Schnürer, Raimund; Sieber, René; Çöltekin, Arzu (2014)
Lecture Notes in Geoinformation and Cartography ~ Modern Trends in Cartography
Atlas der Schweiz – Fit für die Zukunft!
Sieber, René; Schnürer, Raimund (2016)
Geomatik Schweiz
Unsupervised historical map registration by a deformation neural network
Wu, Sidi; Schnürer, Raimund; Heitzler, Magnus; et al. (2022)
GeoAI '22: Proceedings of the 5th ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery
Image registration that aligns multi-temporal or multi-source images is vital for tasks like change detection and image fusion. Thanks to the advance and large-scale practice of modern surveying methods, multi-temporal historical maps can be unlocked and combined to trace object changes in the past, potentially supporting research in environmental science, ecology and urban planning, etc. Even when maps are geo-referenced, the contained geographical features can be misaligned due to surveying, painting, map generalization, and production bias. In our work, we adapt an endto- end unsupervised deformation network that couples rigid and non-rigid transformations to align scanned historical map sheets at different time stamps. To the best of our knowledge, we are the first to use unsupervised deep learning to register map images. We address the sparsity of map features by introducing a loss based on distance fields. When aligning the displaced landmark locations by our proposed method, the results are promising both quantitatively and qualitatively. The generated smooth deformation grid can be applied to vector features directly to align them from the source map sheet to the target map sheet.
Storytelling with Animated Interactive Objects in Real-time 3D Maps
Schnürer, Raimund (2023)
Digital story maps present geographic information embedded in a narrative structure and often supplemented by multimedia elements. The currently predominant extrinsic approach, however, impedes following the storyline since the reader's attention is split between maps and additional illustrative content in the user interface. Incoherent representations between abstract map elements and realistic multimedia elements can be seen as another shortcoming. To close these gaps, an intrinsically oriented approach, which is inspired by historical and contemporary pictorial maps, is proposed in this dissertation: Figurative objects, which act as storytellers, are introduced into the map. By spatially anchoring the complementary entities within 3D maps in particular, the connection between the cartographic model and geographic reality will be strengthened. The overall goal of this work is to automatically turn static objects from prevalent 2D pictorial maps into animated objects for interactive 3D story maps. Artificial neural networks, primarily convolutional neural networks (CNNs), are applied in a sequence of discriminative and generative tasks to achieve the goal. For each task, data is prepared to train the networks in a supervised manner. Firstly, pictorial maps are identified from publicly available images on the internet by CNNs for classification since metadata of the images is not always present or reliable. Different strategies are investigated to input the images into the CNNs. Secondly, bounding boxes of objects on pictorial maps are detected using the example of sailing ships. Although map descriptions may include the occurrences of objects, their positions and sizes are usually unknown. To determine these two measures, CNNs for object detection are examined while modifying their hyperparameters. Thirdly, silhouettes of pictorial objects are recognised, exemplified by human figures. As the manual preparation of training data would be too labour-intensive, combinations of figurative and realistic entities are evaluated. Following, body parts and pose points are extracted from the silhouettes of the figures, which is a prerequisite for skeletal animations and the insertion of speech bubbles at head positions. CNNs with varying numbers of skip connections are compared for this task. Lastly, 3D figures are derived from their 2D counterparts based on the outputs of the previous step by a series of networks. This significantly facilitates and accelerates the 3D modelling process. The figures are represented by implicit surfaces, which are advantageous for curved surfaces, and rendered in real time by a ray tracing algorithm. Quantitative metrics, such as accuracy and rendering speed, and qualitative results are reported for each task. The inferred 3D pictorial objects may guide readers through the map while providing background information and offering interaction possibilities like quizzes. This is especially suitable for atlases, touristic or educative applications, even in augmented and virtual reality. Animated interactive objects are intended to engage map readers through an immersive storytelling approach, increase their map literacy and frequency of use, and create long-lasting memories.
Inferring Implicit 3D Representations from Human Figures on Pictorial Maps
Schnürer, Raimund; Cengiz Öztireli, A.; Heitzler, Magnus; et al. (2021)
Abstracts of the ICA
Human figures frequently occur on pictorial maps besides other illustrative entities. In this work, we present how to automatically derive 3D depictions from these 2D human figures. Previous research has shown that silhouettes, body parts, and joints of 2D human figures in common poses can be detected on pictorial maps by artificial neural networks (Schnürer et al., 2019). Architectures for these networks have been also developed to reconstruct 3D models of real persons from photos in good accuracy (Varol et al., 2018). Single-view methods are particularly suited for our use case since pictorial figures are usually drawn from one perspective only. Furthermore, a trend can be observed to represent the recovered 3D models by implicit surfaces, expressed by level sets of functions (Saito et al., 2019) or signed distance functions (Wang et al., 2019). Compared to other 3D structures, implicit geometries are memory-efficient, but they require special ray tracing algorithms like marching cubes or sphere tracing to be rendered. We examine two approaches: (1) A convolutional neural network, consisting of a feature extractor and a head network, shall learn to directly predict body parts and joints of a 3D model from a 2D image. For this approach, a large amount of training data is essential, for instance, body scans from real persons (e.g. Human3.6M1) or synthetically created persons (e.g. SURREAL2). For our case, these 3D models may be additionally distorted or enriched by rigged human characters from computer games. After converting the geometries from explicit into implicit forms (e.g. mesh-to-sdf3), the network is trained to estimate the resulting values of sample points. (2) Implicit function parameters can be stepwise optimized, for example by Stochastic Gradient Descent, to reduce differences between the target image and its approximation. The latter is a projection of 3D primitives which are combined, transformed, morphed, or deformed by mathematical operations (Pasko et al., 1995). This approach facilitates to formulate constraints such as the connectivity of body parts or rotation angles of joints, but it requires more iterations and eventually ends in a local minimum. The following challenges exist for both approaches: Due to occlusions, multiple reconstruction outputs are plausible. Perhaps, a generative model such as a variational autoencoder or generative adversarial network needs to be introduced to reflect the variety of poses by latent codes. Moreover, a certain strategy may be pursued to sample equally points near the surface, within the body, and in the surrounding space so that local details and thin parts (e.g. fingers) can be preserved (Paschalidou et al., 2020). To speed up the training or optimization process, possibly a meta-learning algorithm may help to find good initialization parameters (Sitzmann et al., 2020). Since human figures on maps are mostly hand-drawn or manually created with graphic software, the camera perspective or lighting conditions may not be fully consistent. It is not clear yet whether this has an impact on differentiable rendering methods (Niemeyer et al., 2020), which may be applied in our networks. Lastly, the texture needs to be mapped to the 3D model and estimated for the hidden parts, which can be achieved by a subnetwork (Saito et al., 2019). We will evaluate the two approaches according to their effectiveness and efficiency. Based on the outcomes of related works and the proposed methods to overcome the challenges, we are optimistic to create meaningful representations. When being successful, the inferred 3D figures could emerge from the original map by augmented reality devices. The figures could then be animated and act as guides on touristic maps or storytellers on historic maps in museums. Due to their attractiveness, the generated 3D figures may raise the interest of people, especially children, in maps and may also serve educative purposes.
Atlas of Switzerland goes online and 3D – Concept, Architecture and Visualization Methods
Sieber, René; Serebryakova, Marianna; Schnürer, Raimund; et al. (2015)
Proceedings of the 1st ICA European Symposium on Cartography
Augmenting Printed School Atlases with Thematic 3D Maps
Schnürer, Raimund; Dind, Cédric; Schalcher, Stefan; et al. (2020)
Abstracts of the ICA ~ Central European Cartographic Conference and 68th German Cartography Congress – EuroCarto 2020
Digitalization in schools requires a rethinking of teaching materials and methods in all subjects. This upheaval also concerns traditional print media, like school atlases used in geography classes. In this work, we examine the cartographic technological feasibility of extending a printed school atlas with digital content by augmented reality (AR). While previous research rather focused on topographic three-dimensional (3D) maps, our prototypical application for Android tablets complements map sheets of the Swiss World Atlas with thematically related data. We follow a natural marker approach using the AR engine Vuforia and the game engine Unity. We compare two workflows to insert geo-data, being correctly aligned with the map images, into the game engine. Next, the imported data are transformed into partly animated 3D visualizations, such as a dot distribution map, curved lines, pie chart billboards, stacked cuboids, extruded bars, and polygons. Additionally, we implemented legends, elements for temporal and thematic navigation, a screen capture function, and a touch-based feature query for the user interface. We evaluated our prototype in a usability experiment, which showed that secondary school students are as effective, interested, and sustainable with printed as with augmented maps when solving geographic tasks.
Augmenting printed school atlases with thematic 3D maps
Schnürer, Raimund; Dind, Cédric; Schalcher, Stefan; et al. (2020)
Multimodal Technologies and Interaction
Digitalization in schools requires a rethinking of teaching materials and methods in all subjects. This upheaval also concerns traditional print media, like school atlases used in geography classes. In this work, we examine the cartographic technological feasibility of extending a printed school atlas with digital content by augmented reality (AR). While previous research rather focused on topographic three-dimensional (3D) maps, our prototypical application for Android tablets complements map sheets of the Swiss World Atlas with thematically related data. We follow a natural marker approach using the AR engine Vuforia and the game engine Unity. We compare two workflows to insert geo-data, being correctly aligned with the map images, into the game engine. Next, the imported data are transformed into partly animated 3D visualizations, such as a dot distribution map, curved lines, pie chart billboards, stacked cuboids, extruded bars, and polygons. Additionally, we implemented legends, elements for temporal and thematic navigation, a screen capture function, and a touch-based feature query for the user interface. We evaluated our prototype in a usability experiment, which showed that secondary school students are as effective, interested, and sustainable with printed as with augmented maps when solving geographic tasks.
Detection of Pictorial Map Objects with Convolutional Neural Networks
Schnürer, Raimund; Sieber, René; Schmid-Lanter, Jost; et al. (2021)
The Cartographic Journal
In this work, realistically drawn objects are identified on digital maps by convolutional neural networks. For the first two experiments, 6200 images were retrieved from Pinterest. While alternating image input options, two binary classifiers based on Xception and InceptionResNetV2 were trained to separate maps and pictorial maps. Results showed that the accuracy is 95-97% to distinguish maps from other images, whereas maps with pictorial objects are correctly classified at rates of 87-92%. For a third experiment, bounding boxes of 3200 sailing ships were annotated in historic maps from different digital libraries. Faster R-CNN and RetinaNet were compared to determine the box coordinates, while adjusting anchor scales and examining configurations for small objects. A resulting average precision of 32% was obtained for Faster R-CNN and of 36% for RetinaNet. Research outcomes are relevant for trawling map images on the Internet and for enhancing the advanced search of digital map catalogues.
The Power of 3D Real-Time Visualization in Atlases – Concepts, Techniques, and Implementation
Sieber, René; Schnürer, Raimund; Eichenberger, Remo; et al. (2013)
Proceedings of the 26th International Cartographic Conference : Dresden, 25-30 August 2013

Publications 1 - 10 of 25

Raimund Schnürer

Last Name

First Name

ORCID

Organisational unit

Filters

Search Results