The ACM SIGSPATIAL Cup 2023 proposed the challenge to identify and map supraglacial lakes in Greenland in satellite imagery. The peculiarities of supraglacial lakes pose a hard problem for semantic segmentation and object detection tasks because the definition of a lake is ill-fitted to the inner workings of such approaches. For example, lakes are often covered by ice and snow and narrow streams can connect distinct lakes, which is not directly translatable to the semantic segmentation of water. It is also not well-posed for object detection, especially the identity relation - what is a lake, what is not (yet) a lake, and what are two lakes is challenging. In this context, we worked on adapting semantic segmentation using the Segment Anything Model and instance segmentation using Mask R-CNN to the setting. The latter ended up superior in our own evaluation and even got ranked second among all participants. We are proud that our approach has led to competitive performance. The source code is available from https://github.com/tum-bgd/GISCup23.
Bavaria Buildings is a large, analysis-ready dataset providing openly available co-registered 40cm aerial imagery of Upper Bavaria paired with building footprint information. The Bavaria Buildings dataset (BBD) contains 18205 orthophotos of 2500 × 2500 pixels, where each pixel covers 40cm × 40cm in space (Digitales Orthophoto 40cm - DOP40). The dataset has been pre-processed and co-registered and also provides a set of 5.5 million image tiles of 250 × 250 pixels ready for deep learning and image analysis tasks. For each image tile, we provide two segmentation masks; one based on the official building footprints (Hausumringe) data as published by the Free State of Bavaria and one based on a historic OpenStreetMap (OSM) extract dating to 2021. The dataset is ready for essential analysis tasks, such as detection, segmentation, instance extraction, footprint geometry extraction, multimodal localization, and multimodal data quality assessment of buildings in Bavaria. We plan to update the dataset with each major re-publication of the upstream data sources to foster change detection research in the future. The BBD is available at https://doi.org/10.14459/2023mp1709451.
Open education is a way of carrying out education, often using digital technologies. This paper gives a brief background on open education and its constraints. To make courses more accessible, teachers need to have an effective way to preserve and share their materials. We derive the requirement of keeping the choice of tools completely open while we pave the way to a teaching transport format based on open standards allowing for extending existing lecture material with interactive content. A simple learning system that helps to detect and save changes of a slide deck was developed. The linkage of relevant knowledge and interactive context can also be added through this implementation which outputs a final transport format. This provides a good starting point for future work on removing learning barriers and widening access to education.
Over time, the peculiarities of point clouds brought forth ample dedicated and specialized solutions for analyzing and managing point cloud data. However, providing analytical capabilities and visualization at scale remains challenging. We present a next-generation point cloud data management approach inspired by the Lakehouse pattern. It is exemplified by combining point clouds stored in raw files with a query engine, which instantly gives us an analysis-ready database management system with an SQL and DataFrame interface.We further demonstrate how to simplify and optimize this system through conversion to a columnar file format and a novel versatile repartitioning approach. Compared to existing solutions, the evaluation exhibits compelling performance, extraordinary flexibility, and exceptional simplicity.
Current point cloud data management systems and formats are heavily specialized and targeted solely towards visualization purposes and fail to address the diverse needs of progressive point cloud workflows like for example semantic segmentation using machine learning. We therefore propose a distributed data infrastructure for dynamic point cloud data management that can support interactive real-time visualization at scale while simultaneously serving as a platform for analytical tasks. By introducing random data distribution, we show that simple query fragmentation and efficient and effective parallelism at scale are possible. At the same time, arbitrary queries in space and time can be efficiently run over the infrastructure including query semantics which returns only a random sample of the query results or preferred points based on an importance dimension calculated, for example, from a local point density information as commonly done in point cloud visualization. To cope with the unknown amount of user-specific attributes and to support even multiple ways of deciding the importance of a given point (ground point removal, coverage of space, random subset) the system is designed to support all of them transparently as multidimensional range queries backed by spatial indices.
Point cloud data analysis and visualization workflows traditionally involve the sequential steps of information retrieval and preceding extensive data preparation. For example, visualizing large point clouds often takes days of processing to translate the data into a suitable representation before visual feedback is possible. While this works fine for static datasets and time-insensitive result consumption, it is unsuitable for dynamic contexts requiring real-time analysis, such as autonomous navigation. To address these shortcomings, we propose a point cloud data management approach based on a log-structured merge-tree that facilitates concurrent and continuous data ingestion and retrieval in real-time at scale. In this paper, we illustrate how to adapt this data structure to the peculiarities of point clouds and how various use cases and query modalities can be supported and optimized by specialized merge operations to repartition and refine the data structure and layout. This includes relying on grid-rounded coordinates and integrating importance as a means for effective and efficient storage, processing, and sampling from point clouds. Initial experiments and evaluation results display promising results and affirm the viability of this approach for Helena, a conceptual next-generation point cloud data management platform for interactive visualization and real-time analytics.