Data science
Table of contents
- F - 24 Literature Review on Learning-Based 3D Reconstruction
- F - 23 Bridge Component Classification from Images
- F - 22 Image Quality Assessment for Online Imagery
- F - 21 Fiber-Optic Measurements Under High-Frequency Load
- F - 19 AI-assisted multi-modal conceptual design for bridges
- F - 1 Damage assessment with ultrasonic measurements
F - 24 Literature Review on Learning-Based 3D Reconstruction
Full title: Learning-Based 3D Reconstruction Using Foundation Models: A Literature Review
Recently, image-based 3D reconstruction approaches have shifted from traditional, incremental Structure from Motion to regression/learning-based approaches. Examples include VGGT, Depth Anything 3, MapAnything, and π³. Many approaches build on top of such foundation models and fine-tune them, in a supervised or unsupervised manner, for specific scenarios.
The objective of this thesis is to conduct a systematic literature review on 3D reconstruction foundation models and on fine-tuning/optimization approaches for such models. Building on this, a method is to be proposed (in theory, not experimentally tested) for how a 3D reconstruction foundation model can be optimized for scenes where the input images differ in acquisition time and are exposed to different environmental conditions (weather, lighting, season). Among others, the following questions are to be addressed:
-
- Which architectures, training data, and capabilities/limitations do current 3D reconstruction foundation models exhibit?
- Which strategies exist for adapting or optimizing these models for specific scenarios?
- How can differences in appearance between images (different acquisition times, weather, and lighting conditions) be handled methodologically?
Possible work steps:
-
- Definition of the search strategy and selection criteria for the literature
- Systematic review of existing 3D reconstruction foundation models
- Systematic review of existing fine-tuning and optimization approaches
- Structured comparison of the approaches (e.g., in the form of a taxonomy or comparison table)
- Derivation of open research questions/gaps
- Conceptual proposal of an optimization method for scenes with varying acquisition conditions, including a suggested evaluation design
The exact task definition will be jointly coordinated and refined before the start and during the work process.
Prerequisites:
Interest in computer vision and deep learning, as well as good English reading skills for scientific literature. Programming skills are not required.
Contact person:
Morris Benedikt Florek
+49 351 46340975
F - 23 Bridge Component Classification from Images
Full title: Classification of Component Properties of Bridges Based on Image Data
Publicly available images store a wealth of valuable information about engineering structures such as bridges. However, this information can only be leveraged if the images are processed automatically. Deep learning-based methods in the field of computer vision enable the use of neural networks for tasks such as classification, object recognition, and semantic segmentation.
Building on top of an existing bridge component object detection dataset that detects the components superstructure, pier, and abutment for girder and frame bridges, this dataset is to be extended by adding per-component classification. Examples include the material, cross-section, or shape (haunched, straight) of the superstructure, as well as the material, shape (e.g., V-shaped), or connection to the superstructure (bearing, monolithic) of the pier.
The objective of this thesis is to develop a suitable attribute catalogue for the components mentioned above, annotate the existing dataset accordingly, and train a model that classifies these attributes based on the detected components. Among others, the following questions are to be addressed:
- Which component properties can be reliably determined visually from images?
- How should the annotation process be structured to obtain consistent and reliable labels?
- How can a classification model be meaningfully trained and evaluated on the detected component crops?
Possible work steps:
- Literature research on fine-grained image classification and multi-label/multi-task approaches
- Development of an attribute catalogue and annotation guideline for each component type
- Annotation of the existing dataset, including a consistency check
- Training and evaluation of a classification model per component and/or attribute
- Analysis of misclassifications and limitations of the approach
The exact task definition will be jointly coordinated and refined before the start and during the work process.
Prerequisites:
Initial experience in programming with Python and common deep learning libraries (e.g., PyTorch), as well as basic knowledge of computer vision. Interest in structural engineering and the use of high-performance computers (HPC) are advantageous.
Contact person:
Morris Benedikt Florek
+49 351 46340975
F - 22 Image Quality Assessment for Online Imagery
Full title: Predicting the Image Quality of Publicly Available Online Imagery
Mapillary is an online platform where users upload street-level images, which are georeferenced and made freely accessible to the public. Mapillary provides an image quality score for its available imagery; however, it is unclear how this score is calculated (documentation: "nullable float, predicted visual quality of the image in the range (0.0, 1.0)"). For information extraction from images sourced from public or online sources, it is important to have access to high-quality images or to filter out low-quality ones, especially for image-based 3D scene reconstruction, but also for classification, object detection, or semantic segmentation.
The objective of this thesis is to build and evaluate a method for predicting the image quality of online imagery. First, the image and acquisition properties that actually drive the Mapillary score are to be investigated, by systematically analyzing scores and their corresponding images under different conditions (e.g., night, rain, overcast, sunshine, blur, occlusions). Building on this, methods for predicting image quality are to be developed, for example (1) using traditional metrics such as resolution, brightness, contrast, or sharpness, and/or (2) by training a small neural network. Both approaches are to be compared against each other and against a ground-truth set sourced from Mapillary.
Possible work steps:
- Literature research on (no-reference) image quality assessment
- Collection of images and corresponding quality scores via the Mapillary API
- Exploratory analysis to identify the factors driving the Mapillary score
- Implementation and evaluation of an approach based on traditional image metrics
- Training and evaluation of a small neural network for quality prediction
- Comparative evaluation of both approaches and discussion of their suitability for filtering online imagery
The exact task definition will be jointly coordinated and refined before the start and during the work process.
Prerequisites:
Interest or initial experience in programming with Python and common libraries (e.g., NumPy, Pandas, OpenCV, PyTorch) as well as working with REST APIs are advantageous.
Contact person:
Morris Benedikt Florek
+49 351 46340975
F - 21 Fiber-Optic Measurements Under High-Frequency Load
Ensuring the safety of aging infrastructure is of great importance. Distributed fiber-optic sensors (DFOS) are increasingly used for early damage detection, as they continuously measure strain over extended lengths with high spatial resolution, thereby offering the potential to detect cracks in concrete structures. However, challenges arise with high-frequency strain changes, such as those caused by traffic loads on bridges. These can impair the quality of the measurement data and thus the reliability of crack detection.
As part of this project, measurement data from experiments on the openLAB research bridge will be used to evaluate the potential of DFOS for strain measurement and crack detection under dynamic loading. The work consists of the following tasks:
- Literature review on distributed fiber-optic sensors with a focus on measurements under dynamic loading
- Analysis of experimental data to investigate the influence of high-frequency loads on the availability and quality of measurement data, as well as on the reliability of crack width calculations
- Proposal of solutions to improve the measurement data quality of DFOS under dynamic loading
Details of the task will be refined prior and while working on the project. Interest/experience in software development/programming is advantageous.
Contact:
Miriam Kroschel
+49 351 463-36073
F - 19 AI-assisted multi-modal conceptual design for bridges
In the early stages of bridge design (conceptual design), engineers must quickly generate and evaluate numerous bridge alternatives to satisfy diverse terrain, environmental, structural and economic requirements. Traditional manual methods are typically time-consuming, resource-intensive, and rely heavily on designers' experiences, limiting the systematic exploration of optimal design solutions. The integration of Artificial Intelligence (AI) with multi-modal data, including numerical parameters, geometric data, textual standards and historical design data, could significantly enhance decision-making, accelerate the generation of feasible design concepts and improve early-stage evaluations regarding cost, feasibility and sustainability.
The primary objective of this research is to develop an AI-assisted multi-modal framework capable of efficiently generating and evaluating conceptual bridge designs, specifically for slab, slab-beam, and composite bridges. The approach aims to automate preliminary bridge type selection, estimation of key cross-sectional parameters, and rapid evaluation of multiple early-stage design options in terms of structural feasibility, cost-efficiency, and sustainability. Based on this objective, the following tasks are involved:
- Literature research regarding potential AI algorithms for the engineering design problem
- Collect and structure multi-modal data (project parameters, site geometry, historical designs, and design guidelines)
- Develop AI models for automatic bridge type selection and preliminary prediction of cross-sectional and geometric parameters
- Implement a rule-based validation tool to ensure conceptual designs meet structural and regulatory requirements
- Perform preliminary evaluations of material usage, construction costs, and carbon emissions for design comparison
The work is part of the research project mFUND-HyBridGen – Hybrid Bridge Generator: AI-based bridge generator with knowledge and experience data and early citizen participation. Details of the task will be refined prior and while working on the project. Interest/experience in AI-based methods for structural engineering (civil/computational engineering) is advantageous.
Contact:
Han Qian
+49 351 463 33083
F - 1 Damage assessment with ultrasonic measurements
Full title: Damage assessment of cyclically loaded concrete structures with ultrasonic measurements
Concrete structures under a given load do not fail because they abruptly change from a "normal" state to a fracture state, but because the degradation process progresses with increasing load until material failure occurs. When subjected to mechanical loads, stresses first concentrate around material defects or interfaces at the microscale, destroying bonds between individual molecules. With increasing mechanical load, the microcracks then grow and unite, leading to the formation of macrocracks. During this process, the lattice structure of the material, which serves as a propagation medium for the stress waves of an ultrasonic pulse, is progressively changed and in this way the damage can be detected.
The objective of this thesis is to relate ultrasonic measurements of degradation evolution from concrete specimens and beams subjected to cyclic loading to hypotheses of damage accumulation. From these correlations and using concepts of robustness and redundancy, safety factors will be determined and the remaining useful life will be evaluated.
Contact:
Raúl Enrique Beltrán Gutiérrez
0351 463 33675