Possible Topics
| Topic | Short description | Type of degree | Contact person |
|---|---|---|---|
| Segmentation of rare classes using foundation models on 2D images |
Segmentation of areas and detection of objects is one of the most fundamental tasks in the field of remote sensing and image analysis. Machine learning models that have been trained on a training data set are often used for this purpose. The typical classes of a training dataset are: buildings, roads, forests, agriculture, …. But what happens with not-before seen areas and objects? In the case of natural disasters, these models would struggle with destroyed buildings, flooded areas, muddied roads or burnt forests. Working on imageries of natural disasters, like the Ahrtal flooding in 2021, these unusual areas need to be identified using an already trained machine learning model. That model can be personalized to new objects using tools already available. Still, some knowledge of programming in python and Matlab is recommended. |
Master | Kevin Qiu and Rewanth Ravindran , Co-supervised by Fraunhofer IOSB |
| Tree parameter extraction using LiDAR 3D point cloud data | Machine Learning methods are really transforming what can be done using remote sensing images in various environmental applications – from monitoring deforestation to identifying cracks in polar ice sheets. With the advent of drone based Laser-Scanning (LiDAR) data, the quality of details in the data in urban or agriculture or forest contexts is higher than ever before. 3D point clouds from LiDAR and machine learning models provide a very powerful combination for extracting useful information such as tree canopy segmentation or estimating biomass of the woody stems and branches, with relatively high accuracy. Before the model can be used for its individual purpose, it must be trained by feeding it with labelled training data. Generating training data by manually mapping hundreds of canopies in aerial photographs, for example, is time-consuming, labour-intensive and annoying. But with lack of sufficient labelled training dataset to create a new model, what other options exist? Can we use already existing general-purpose models built for image segmentation (see SAM by Meta AI) for our applications? Can we minimally adapt these models using existing methods so that we produce better accuracy in estimating tree parameters? These are some of the questions that we aim to be addressed by the master thesis. | Master | Rewanth Ravindran and Kevin Qiu , Co-supervised by Fraunhofer IOSB |
| Your own topic proposal on the subject of 3D point clouds, laser scanning, mobile mapping | If you are interested in working with data from 3D sensor systems but the presented topics to not fully fit your interests/requirements/qualification, just go ahead and contact us. In most cases, we will be able to find a topic for you fitting your specific interests and skills, both on bachelor and master level. Currently, we are mostly focusing on capture, automated reconstruction and all kinds of analysis of 2 specific kinds of environments: indoor building environments and forests. But we are also flexible towards other kinds of application scenarios if you come along with a specific idea. | Bachelor / Master | Mona Goebel Patrick Hübner |
| Classification of scanned building fragments for the reconstruction of historical buildings | The fragments of a Jewish Bima were captured by laser scanning. From these fragments the whole building was reconstructed manually. In this work, point clouds of the laser scanned fragments will be classified as building parts using neural networks (e.g. PointNet++). This serves as a first step to simplify the puzzling together of the fragments. | Master | Jakob Schmidt |
|
Voxel-based Indoor Building Model Reconstruction with Machine Learning Methods |
Scan2BIM, i.e. the automated reconstruction of building models from 3D sensor data like mobile mapping point clouds is a very active field of research driven by the increasing demand in building industry, architecture and facility management for accurate and up-to-date BIM models for existing buildings. The research group has an own, voxel-based approach for reconstructing indoor building environments in the form of 3D voxel grids (More). While the method has proven to be applicable to a wide range of different building shapes, it still relies on a set of hand-crafted rules. Data-driven methods based on machine learning hold greater potential for more flexibility. Thus, we want to try out different methods from classical machine learning and deep learning in comparison to our rule-based baseline approach. | Bachelor / Master | Patrick Hübner |
|
Completion of Hidden Wall Surfaces on Existing Datasets of Indoor Building Environments with Semantics |
Scan2BIM, i.e. the automated reconstruction of building models from 3D sensor data like mobile mapping point clouds is a very active field of research driven by the increasing demand in building industry, architecture and facility management for accurate and up-to-date BIM models for existing buildings. While meanwhile large-scale 3D indoor mapping datasets with thousand of buildings exist, these dataset provide ground truth typically only in the form of semantic labeling of the represented geometries, i.e. the information for each point/triangle if it belongs to a wall, stair, furniture etc. Ground truth information on hidden building surfaces which are not represented in the indoor mapping geometries (e.g. hidden by furniture) however, is rarely available but would be of great importance for training learning-based Scan2BIM approaches. We want to make use of the fact that in the existing datasets, we at least have semantics and thus can easily identify furniture objects excluding wall surfaces. On this basis, we want to work towards enriching the existing large-scale indoor datasets by surface completion in an automated or semi-automated manner. | Bachelor / Master | Patrick Hübner |
|
Image-Based Indoor Building Model Reconstruction |
Scan2BIM, i.e. the automated reconstruction of building models from 3D sensor data like mobile mapping point clouds is a very active field of research driven by the increasing demand in building industry, architecture and facility management for accurate and up-to-date BIM models for existing buildings. A broad range of approaches for reconstructing builidng models based on 3D sensor data such as laser scanners already exists. Active sensor system however are costly and not as commoly available as cameras. Due to the specific challanges of indoor environments, approaches for reconstructing indoor building models purely based on image data are still rare. Current progress in photogrammetry, e.g. by incorporating deep learning into the pipeline or by using implicit neural representations instead of explicit point geometries as scene representations offer new possibilities to deal with this challenging problem. We want to investigate this field of research in a series of bachelor and master theses. |
Bachelor / Master | Patrick Hübner |
|
Large-Scale Automated Evaluation of SLAM Algorithms in Virtual Environments |
SLAM algorithms are used to automatically determine sensor trajectories and 3D maps of environments a mobile sensor system is moving in based on image sequences or sequences of laser scanning data. Evaluating the quality of SLAM algorithms however is laborious and mostly done in limited setting, i.e. on a low number of datasets and trajectories. In recent years large-scale interactive, virtual photo-realistic 3D enviroments comprising thousands of different buildings are being made available as training enviroments for self-learning autonomous robots (e.g. Gibson Environment, Habitat Matterport 3D). We as surveyors see these virtual worlds as a chance for precise and detailed, fully-automatic quantitative evaluation of different kinds of SLAM algorithms that will allow us to assess their quality on thousands of different trajectories in different environments and want to work towards this goal in a series of bachelor and master theses. |
Bachelor / Master | Patrick Hübner |
|
Fusing Close-Range Multispectral Data and Geometric Reconstruction |
Multispectral cameras provide a wide range of valuable information about vegetation health. They are however typically deployed on airbone platforms such as planes and drones and used for large-scale aerial image analysis. We are interested in applying these sensors in closer-range scenarios such as analysing in-forest scenarios from ground-based mobile mapping systems on the level of single plants. In this context, the complicated, fine-grained 3D geometry of vegetation plays are larger role compared to bird-eye scenarios of drone-based applications. In this context, we want to investigate how to fuse ground-based multispectral image data with geometric data from other sensors such as lidars or depth cameras. One family of approaches that are of great interest in this context are neural radiance fields (NeRF). |
Bachelor / Master | Patrick Hübner |