[{"data":1,"prerenderedAt":154},["ShallowReactive",2],{"jobs-list-de":3,"thesis-list-de":4},[],[5,30,48,65,82,99,116,136],{"id":6,"title":7,"author":8,"body":9,"date":20,"description":15,"extension":21,"meta":22,"navigation":23,"path":24,"seo":25,"status":26,"stem":27,"type":28,"__hash__":29},"thesis\u002Fde\u002Fthesis\u002F20260612_guezguez.md","Evidential Fusion of Frozen Heterogeneous Backbones for Reliable FewShot Out-of-Domain Classification","Mohammed Amine GuezGuez",{"type":10,"value":11,"toc":16},"minimark",[12],[13,14,15],"p",{},"Deploying an image classifier on an unfamiliar domain from a handful of labels\ndemands two things: accuracy, and a reliable signal of when it is likely wrong.\nFrozen pretrained backbones supply the accuracy. A single backbone, though, can\nfalter where its inductive bias does not suit the domain, and one that cannot flag\nsuch failures is hard to deploy. This thesis asks how several frozen, heterogeneous\nbackbones should be combined to meet both demands. Starting from a gated\nmulti-backbone model, it first locates the gain. It comes from combining the per-\nbackbone decisions, not the learned gating or the feature-level fusion, both inert\nhere.\nThe thesis then develops an evidential fusion of the backbones. It combines their\nper-backbone Dirichlet opinions through the reduced Dempster rule, a parameter-\nfree successor to the learned gate that trains only a lightweight head per backbone.\nOn the original ImageNet-trained backbones a single strong modern backbone still\nbeats the combination, showing that diversity is necessary but not sufficient: back-\nbone strength matters as much as complementarity. A pre-hoc selection on cost\nand diversity, measured by representation similarity and prediction disagreement,\nthen pairs two complementary, comparably strong backbones, DINOv2 ViT-B\u002F14\nand ConvNeXt-V2 Tiny.\nAcross four out-of-domain datasets and one to twenty labels per class, the combi-\nnation has no weak dataset. It leads on accuracy, keeps a consistently low selective\nrisk, is among the best calibrated after the same temperature scaling is applied to\nevery method, and stays competitive on misclassification detection. The single-\nbackbone baselines, including the evidential and Monte-Carlo dropout variants,\ninstead inherit the weakness of the backbone beneath them. Its evidential vacuity\ngives a trustworthy epistemic signal on all four domains, where that of the single\nevidential models is erratic, though entropy remains the sharper error detector.\nThe cost is modest: about 1.3 times the parameters and 1.5 times the latency of a\nsingle strong backbone, no more than a deep ensemble of the same pair. The value\nof the combination therefore lies less in a large accuracy margin than in equally\naccurate but more consistent and more trustworthy predictions. That holds only\nwhen its backbones are complementary and comparably strong.",{"title":17,"searchDepth":18,"depth":18,"links":19},"",2,[],"2026-06-12","md",{},true,"\u002Fde\u002Fthesis\u002F20260612_guezguez",{"title":7,"description":15},"closed","de\u002Fthesis\u002F20260612_guezguez","bachelor","1Q3zsQP2zd2OXGo52kmJ3ozCTOKeEb8RAByUjH2JxKk",{"id":31,"title":32,"author":33,"body":34,"date":41,"description":38,"extension":21,"meta":42,"navigation":23,"path":43,"seo":44,"status":26,"stem":45,"type":46,"__hash__":47},"thesis\u002Fde\u002Fthesis\u002F20251001_schwass.md","Multi-Backbone Embedding-Fusion: Ein attention-basierter Deep-Learning-Ansatz für die Fusion heterogener Computer-Vision-Backbones","Jacob Schwaß",{"type":10,"value":35,"toc":39},[36],[13,37,38],{},"Diese Arbeit untersucht die Fusion latenter Repräsentationen mehrerer heterogener Computer-Vision-Backbones als Ansatz zur gezielten Strukturierung und Stabilisierung des resultierenden Embedding-Spaces. Ziel ist die Analyse der Bedingungen, unter denen lernbasierte Fusion zu trennschärferen, semantisch konsistenteren und daten-effizienteren Repräsentationen führt. Hierzu werden fünf vortrainierte Backbone-Architekturen als reine Feature-Extraktoren eingesetzt. Die erzeugten Embeddings werden mittels lernbarer Adapter in einen gemeinsamen latenten Raum projiziert und durch ein attention-basiertes Fusionsmodell aus Multi-Head-Self-Attention und additiver Attention zu einer gemeinsamen Repräsentation integriert. Die Evaluation erfolgt auf ImageNet100 sowie auf einem domänenspezifischen Datensatz zur Jakobskreuzkraut-Erkennung und umfasst Klassifikations-, Struktur-, Robustheits- und Laufzeitanalysen sowie Few-Shot-Experimente. Ergänzend werden Visualisierungen und Interpretierbarkeitsanalysen mittels UMAP, Korrelationsuntersuchungen der Backbone-Embeddings und Gradient × Input durchgeführt. Die Ergebnisse zeigen, dass lernbasierte Fusionsstrategien insbesondere bei moderater oder heterogener Repräsentationsqualität der zugrunde liegenden Backbones deutliche Zugewinne erzielen. Die Fusion verbessert Clusterkohärenz und globale Struktur des Embedding-Spaces, während klassische Verfahren wie Mittelung strukturelle Grenzen aufweisen. Einfache Konkatenation erreicht teilweise vergleichbare Performanz, jedoch auf Kosten stark wachsender Dimensionalität und ohne explizite Informationsintegration oder -verdichtung. Unter Domänenverschiebungen stabilisiert die Fusion bestehende Merkmalsräume, erweitert deren Generalisierungsfähigkeit jedoch nicht grundlegend. Im Few-Shot-Regime ermöglicht sie eine effizientere Nutzung begrenzter Annotationen, wobei der Vorteil mit zunehmender Datenmenge abnimmt. Insgesamt zeigt die Arbeit, dass Embedding-Fusion keinen generellen Ersatz für leistungsstarke Einzelbackbones darstellt, unter geeigneten Bedingungen jedoch ein wirkungsvolles Verfahren zur strukturellen Integration heterogener Repräsentationen in praxisnahen Computer-Vision-Anwendungen bietet.",{"title":17,"searchDepth":18,"depth":18,"links":40},[],"2025-10-01",{},"\u002Fde\u002Fthesis\u002F20251001_schwass",{"title":32,"description":38},"de\u002Fthesis\u002F20251001_schwass","master","1otK0QMu80Qkm0oczNmrYW4_oDQmt1jtfBUkVSAYo6U",{"id":49,"title":50,"author":51,"body":52,"date":59,"description":56,"extension":21,"meta":60,"navigation":23,"path":61,"seo":62,"status":26,"stem":63,"type":46,"__hash__":64},"thesis\u002Fde\u002Fthesis\u002F20250117_lehn.md","Developing a Domain-Specific Foundation Model with Adapters and Self-Supervised Learning for Efficient Plant Detection in Grasslands","Mirko Lehn",{"type":10,"value":53,"toc":57},[54],[13,55,56],{},"The rise of foundation models has enabled generalizable solutions across diverse computer vision tasks. However, their performance often declines in specialized domains where data is limited, annotations are sparse, and subtle visual cues dominate. This thesis targets such a scenario: detecting individual plants in grassland fields where low inter-class variance, weak contrast, and small objectsizes pose persistent challenges. Instead of discarding prior work like DETReg, which was criticized for its lack of classification and reliance on generic object priors, this thesis revisits and extends it. Building on the Deformable Detection Transformer (Deformable DETR), the proposed method introduces an unsupervised, domainaware pretraining pipeline that directly addresses DETReg’s limitations. It uses custom pseudo boxes tailored to the target domain and gradually incorporates teacher model predictions via an Exponential Moving Average (EMA)-based training loop. Semantic learning is introduced early through a binary classification signal, and regularization mechanisms are applied to stabilize training and avoid collapse. While results show only marginal gains when abundant labeled data is available, the benefits become pronounced under few-shot conditions. The proposed pretraining strategy consistently outperforms models initialized from unrelated domains such as COCO. Moreover, it is designed to operate with minimal supervision and for reducing not only labeled data requirements but also the need for large volumes of unlabeled data by leveraging curriculum schedules and early semantic cues. Lightweight adapter modules are optionally integrated to stabilize the training in the early phase and to test additional possibilities for data reduction, but are not decisive for the final performance. This work presents a scalable and interpretable approach to domain-specific pre-training, aimed at data efficiency and practical applicability in real world, lowresource scenarios. Applications include ecological monitoring, automated weed control, and other tasks requiring finegrained recognition in complex natural environments.",{"title":17,"searchDepth":18,"depth":18,"links":58},[],"2025-01-17",{},"\u002Fde\u002Fthesis\u002F20250117_lehn",{"title":50,"description":56},"de\u002Fthesis\u002F20250117_lehn","KabB9yB_qgUO7zvbhdZnecrfR_2eSKnHriM8G9U6CBY",{"id":66,"title":67,"author":68,"body":69,"date":76,"description":73,"extension":21,"meta":77,"navigation":23,"path":78,"seo":79,"status":26,"stem":80,"type":46,"__hash__":81},"thesis\u002Fde\u002Fthesis\u002F20241224_reuter.md","Semantic Interoperability in the Web of Things: Design and Validation of a Reference Data Model for Monitoring Applications in the Grasslands Agricultural Domain","Sebastian Reuter",{"type":10,"value":70,"toc":74},[71],[13,72,73],{},"There currently exists no ontology for the pest management subdomain of precision agriculture. This prevents the development of interoperable Internet of Things (IoT) solutions, and cements the existing heterogeneous landscape of vertical solutions. In the context of developing a weed detection device for KI-gestützte hochautomatisierte Unkrautbekämpfung im Grünland (KIhUG), a research project at the Technische Hochschule Mittelhessen (THM) a partial ontology has been designed with the goal of making the device highly interoperable with third-party weed management systems. The ontology and the accompanying set of semantic data models proved to be a valid starting point for the design of a full-scale pest management ontology.",{"title":17,"searchDepth":18,"depth":18,"links":75},[],"2024-12-24",{},"\u002Fde\u002Fthesis\u002F20241224_reuter",{"title":67,"description":73},"de\u002Fthesis\u002F20241224_reuter","tn8MLWdMBvKEr_cOxfmAhx1v7-_yOsg_oh440pYprLw",{"id":83,"title":84,"author":85,"body":86,"date":93,"description":90,"extension":21,"meta":94,"navigation":23,"path":95,"seo":96,"status":26,"stem":97,"type":46,"__hash__":98},"thesis\u002Fde\u002Fthesis\u002F20240409_vaupel.md","Generation of Synthetic Training Data for Neural Network based Weed Detection with Stable Diffusion","Dennis Vaupel",{"type":10,"value":87,"toc":91},[88],[13,89,90],{},"Artificial intelligence is becoming a key technology in many industries, including agriculture. Precision farming heavily leverages modern computer vision technology to equip farm machines as well as autonomous robotic platforms with perceptive capabilities. For years, neural networks have been the first choice when it comes to machine vision. Although robust and high-performing, neural networks come with a caveat: Their training requires a great amount of annotated training data. This presents an obstacle to machine learning engineers who need a certain dataset size to reach their required performance metrics. Collecting data themselves is often not economically feasible for agriculture companies. The annotation process is also labor-intensive and takes valuable time. Synthetic training data provide a remedy to this concern. This work develops a framework to generate synthetic image data along with pixel-wise annotations. It presents a multi-stage ensemble reference architecture and demonstrates its potential with an example of weed in grassland. The work concludes with a critical discussion and possible real-world applications.",{"title":17,"searchDepth":18,"depth":18,"links":92},[],"2024-04-09",{},"\u002Fde\u002Fthesis\u002F20240409_vaupel",{"title":84,"description":90},"de\u002Fthesis\u002F20240409_vaupel","cU7mDP-PfO-B23IiPWFaDOIBP96-obhN5TBFW-XMdoA",{"id":100,"title":101,"author":102,"body":103,"date":110,"description":107,"extension":21,"meta":111,"navigation":23,"path":112,"seo":113,"status":26,"stem":114,"type":46,"__hash__":115},"thesis\u002Fde\u002Fthesis\u002F20240301_hohl.md","Deep Learning based Growth Modeling of Plant Phenotypes","Renke Hohl",{"type":10,"value":104,"toc":108},[105],[13,106,107],{},"This master’s thesis investigates the largely unexplored domain of image-based plant growth modeling using deep learning, aiming to generate synthetic training data for object detection and segmentation across different plant growth stages. Demonstrating the ability of deep learning models to simulate growth of plants this study introduces a novel class of neural network architectures called Latent Growth Models (LGMs), which attempt to model the temporal aspects of plant growth exclusively within the efficient latent space of pre-trained autoencoders. Specifically, the research of this thesis includes the development and investigation of three distinct LGM architectures, some of which demonstrate superior performance over existing state-of-the-art methods in terms of image quality, while retaining a high degree of realism in the simulated growth, leveraging the Aberystwyth Leaf Evaluation dataset for empirical validation. Furthermore, the architectures achieve comparably fast training times, enhancing their practical applicability and setting a new benchmark in the field. Despite these advancements, this thesis identifies a significant gap in the current research landscape: the absence of specialized metrics capable of evaluating the realism with which plant growth is simulated in image data. This gap highlights a crucial direction for future research, emphasizing the development of robust evaluation tools to quantitatively assess the authenticity and applicability of synthetic image data in advancing the field of plant phenotyping and beyond. Furthermore, the thesis delineates strategies for refining the proposed methodologies in subsequent studies. These primarly consider architectural details and training strategies of the models, with a particular focus on enhancing image-based modeling of plant growth.",{"title":17,"searchDepth":18,"depth":18,"links":109},[],"2024-03-01",{},"\u002Fde\u002Fthesis\u002F20240301_hohl",{"title":101,"description":107},"de\u002Fthesis\u002F20240301_hohl","oXLMEuR01QNtaxVN0UR6YN2-u8Im5uGcILPNq8JOBBo",{"id":117,"title":118,"author":119,"body":120,"date":130,"description":124,"extension":21,"meta":131,"navigation":23,"path":132,"seo":133,"status":26,"stem":134,"type":28,"__hash__":135},"thesis\u002Fde\u002Fthesis\u002F20231201_koch.md","Generierung synthetischer modellbasierter Bilddaten zum Training von Künstlichen Neuronalen Netzen","Till Koch",{"type":10,"value":121,"toc":128},[122,125],[13,123,124],{},"Um Künstliche Intelligenz (KI) zur Objektdetektion zu trainieren, werden große Mengen annotierter Daten benötigt. Mit steigendem Bewusstsein über KI eröffnen sich auch zunehmend vielfältige Anwendungsgebiete. Ein exemplarisches Einsatzfeld ist die Landwirtschaft, insbesondere im Bereich des Unkrautjätens. Doch gerade die Erstellung spezifischer Datensätze, beispielsweise zur Unkrautdetektion, kann zeitaufwändig und kostenintensiv sein.",[13,126,127],{},"Um diesen Prozess zu erleichtern, werden in dieser Bachelorthesis synthetische modellbasierte Bilddaten von Weideland zum Training von künstlichen neuronalen Netzen generiert. Dabei wird sowohl eine Pipeline zum Erstellen dieser Daten beschrieben, als auch deren Anwendbarkeit auf realen Aufnahmen getestet.",{"title":17,"searchDepth":18,"depth":18,"links":129},[],"2023-12-01",{},"\u002Fde\u002Fthesis\u002F20231201_koch",{"title":118,"description":124},"de\u002Fthesis\u002F20231201_koch","CqZDlmV4EZhFrM90QAaUhwYRyVYJ-FF56dO387Q_syM",{"id":137,"title":138,"author":139,"body":140,"date":147,"description":144,"extension":21,"meta":148,"navigation":23,"path":149,"seo":150,"status":26,"stem":151,"type":152,"__hash__":153},"thesis\u002Fde\u002Fthesis\u002F20231128_soethe.md","Evaluierung und Auswahl eines Objekterkennungsmodell zur Unkrautdetektion im Grünland","Nico Soethe",{"type":10,"value":141,"toc":145},[142],[13,143,144],{},"Diese Bachelorarbeit beschäftigt sich mit der Objekterkennung von Jakobskreuzkraut (JKK) im Grünland. Das Ziel der Arbeit ist es, einen Detektor  zu erstellen, der unter variablen Umwelteinflüssen genutzt werden kann. Ein zuverlässiger Detektor ist essenziell, um Landwirte beim Erkennen und potenziell automatisierten Entfernen dieser giftigen Pflanze zu unterstützen, welche in der Heuproduktion gravierende Gefahren birgt. Nach der Erstellung des Datensatzes wurden mehrere neuronale Netze trainiert und anhand von vorgegebenen Anforderungen wie Erkennungsgenauigkeit und Verarbeitungsgeschwindigkeit evaluiert. Abschließend wurde ein Modell ausgewählt, das die besten Ergebnisse im Vergleich zu anderen Modellen und experimentellen Setups erzielt hat.",{"title":17,"searchDepth":18,"depth":18,"links":146},[],"2023-11-28",{},"\u002Fde\u002Fthesis\u002F20231128_soethe",{"title":138,"description":144},"de\u002Fthesis\u002F20231128_soethe","Bachelor","kbggaMqM8rdrmJfE8gCqD7e9o42px7-0zn9Rt6bncLs",1793017187204]