Plant phenotyping plays a vital role in plant genetics and breeding programs, providing the foundation for screening and evaluating genetic diversity and linking phenotypic parameters to the genetic determinants of trait expression. This process is critical for identifying molecular markers and accelerating genetic breeding improvement, thereby enhancing plant resilience to biotic and abiotic stresses such as drought, salinity, and diseases. Recent advancements in 3D sensing technology have empowered researchers to extract precise phenotypic parameters from plant point clouds, enabling more detailed and accurate plant phenotyping. A critical step in point cloud-based plant phenotyping is plant organ segmentation. Among the available segmentation methods, skeleton-based approaches are simple and intuitive, and could leveraging both local and global geometric information from plant point clouds to facilitate accurate organ segmentation. These methods have gained considerable attention in recent years. However, plant skeletonization, the core component of these approaches, remains limited in handling leafy plants, especially herbaceous species with complex shoot architectures, lateral stems, and multiple leaves. These structural complexities pose challenges that current skeletonization techniques struggle to address effectively.

To address this challenge, we propose a skeleton-based plant organ segmentation framework that accurately extracts curve skeletons from individual plant point clouds and performs precise stem-leaf segmentation based on these skeletons. Our framework is particularly effective in handling leafy plants. It preserves fine structural details during skeletonization while avoiding abnormal or noisy local branches by extending the Laplacian-based contraction (LBC) algorithm through the integration of the Constrained Laplacian Operator. Moreover, we introduce Adaptive Constraints and Tip Points Preservation within the contraction loops to further refine skeleton quality. Additionally, a modified Locally Optimal Projection (LOP) operator is utilized to perform skeleton points calibration, ensuring that the extracted skeleton is centrally aligned with the original plant shape. Furthermore, to evaluate the performance of our proposed framework, we contribute a photogrammetric 3D plant point cloud dataset of 56 Polygonum lapathifolium plants, complete with detailed annotations. Experiment results demonstrate that our framework robustly handles various shapes and sizes of leafy plants and tree branches.

In conclusion, our study enhances the LBC algorithm by integrating the Constrained Laplacian Operator, Adaptive Constraints, and Tip Points Preservation. These improvements increase the accuracy and quality of curve skeleton extraction from leafy plant point clouds, enabling satisfactory plant organ segmentation.

Style transfer is a recent field in the development of deep neural networks, which allows for the style from one image to be transferred onto another image. This has been well-researched for 2D images, but transferring style onto 3D reconstructed content can still be further developed. Being able to style a 3D reconstruction would allow users to recreate anything in the real world, such as a chair, with any style they see fit. Where other methods use texture-based approaches which often create low quality geometry and appearance, or
use radiance fields which style a whole scene instead of just the 3D reconstructed object, we have developed a method which styles an implicit surface.
We achieve this by using Implicit Differentiable Renderer (IDR), which trains, using masked
images as input, two neural networks that learn the geometry and appearance. Rendered views of the object are styled using 2D neural style transfer (NST) methods, and the style information is used to further train the appearance network to display the given style. With Masked deferred back-propagation we are able to optimize the appearance renderer, which is normally trained on only patches of the rendered image to save memory, while using style
transfers designed for full-resolution images. We showcase different results from our method using different 3D reconstruction datasets and style images, and showcase how to implement a user-created dataset. We carry out
extensive tests on what effects different parameters have on the final result. Comparing our results to similar 3D style methods demonstrates that our method performs equally well in achieving faithful style transfer, while having the benefits of creating high quality geometry and only styling the reconstructed surface.

Creating GPS art on the map is an interesting way to make one’s outdoor activity more engaging. Cyclists, runners and hikers can create impressive drawings on the map by traversing the road/pedestrian network in a carefully planned way. Such planning, however, is often tedious and time consuming, which makes the GPS artists have to meticulously design the routes with the complex road network in mind. The aim of our research is to come up with a full process that can express a person’s initial idea (for example a contour drawing) as a route, which the user can then follow to create their GPS art. This involves transforming the input image to match the routing network in a selected area and generating a route which approximates the shape in the best possible way. In our work, searching for patterns in the road network is cast to an image matching problem with template matching as the solution. Generating routes is achieved using a graph routing algorithm with a custom cost function, to make the resulting route as similar to the input shape as possible. Finally, two ways of generating artistic routes are presented. First is an automatic GPS art workflow, which attempts to find an optimal initial location of the route, then generates a number of candidate routes and selects the best one according to various evaluation criteria. The second method is an interactive browser application, where the user can select an initial location for his shape on the map, move, scale or rotate it and get instant feedback in the form of artistic routes displayed in real time.

The reconstruction of 3D city models has garnered significant interest in recent years. However, the majority of existing reconstruction methods primarily focus on LOD2 models, while LOD3 model reconstruction often relies on manual labor, and the primary data sources are street view images. This research aims to advance this field by reconstructing LOD3 models through the addition of windows and doors to existing LOD2 models, thereby maximizing the utility of available 3D building models, as well as the accurate addition of windows and doors. This research innovatively utilizes aerial oblique images as the data source for extracting building openings and employs 3D BAG LOD2.2 models as the basic 3D building structures. The 3D facades are projected onto the 2D aerial image space using perspective projection and registration is employed on the projection facade and oblique aerial images. Subsequently, Mask R-CNN is employed to detect and extract the building openings from these projections. Following the extraction, the layout of the openings within the same facade is optimized in terms of both size and position. Lastly, the relative positions of the openings on the facade images are combined with the 3D coordinates of the corresponding facade to calculate the positions of the openings in 3D space. This information is then integrated into the LOD3 model, resulting in a more detailed and accurate representation of the buildings.
This approach successfully reconstructs the final LOD3 model in CityJSON format, which passes the val3dity validation. By effectively utilizing existing 3D building models, this approach conserves a considerable amount of computational resources required for reconstruction. The simplicity and high level of automation of this approach make it a promising solution for reconstructing large-scale LOD3 buildings, leading to more accurate and detailed large 3D urban models.

3D building models play an important role in many real-world applications. Different models are suitable for different application scenarios based on their levels of detail. LOD3 models with facade details are crucial for many applications, such as virtual reality and urban simulation. Currently, 3D building models with lower LOD are largely available, but the number of LOD3 models is very limited. Most LOD3 reconstruction methods depend on manual operation, which is very time-consuming. How to automatically reconstruct the detailed facade for building models has remained a problem in computer vision. The problem can be seen as an image processing problem, but how to convert the 2D results into 3D smoothly should also be considered. 

In this project, we proposed a method to automatically reconstruct the detailed building models based on the Faster R-CNN. The method starts from a set of street view images, and the results are models with facade elements. A 3D point cloud can be extracted from the images using SfM and MVS, and the camera parameters can also be recovered. We take advantage of the high-quality facade images and parse the facades to detect their bounding boxes. The bounding boxes can match pretty well with the rectangular shape of the facade elements. The 2D facade elements can be added to the 3D building model based on the camera parameters. The process is very efficient and automatic. The regularity of the facade elements will be reserved, making the result more convincing. Our method includes four main steps:  (1) coarse model reconstruction, (2) facade image selection and rectification, (3) facade element detection and regularization, and (4) detailed facade reconstruction.

Experiment results show that our method can produce reliable building models with facade details for many different situations. It can work for both the multi-face building blocks and the street side buildings. Our test shows that the window detection performance is pretty good. The object detection is extremely fast, and the whole pipeline is lightweight and efficient. In theory, the method can also be extended to reconstruct large-scale city models, which means it has broad application prospects.

A digital reconstruction of real-life trees could provide many benefits in fields such as botany, forestry management, biology, and urban planning. Plant growth modelling in particular would enable the analysis of plant structure and behaviour in a customizable, widely applicable and non-destructive manner. Although many data-driven plant reconstruction methods exist, it remains a complex problem due to the intricate branching systems of trees and the need for balancing model soundness with adherence to the often incomplete input data. Modelling plant growth currently proves difficult as well due to the large number of factors involved in the growth process and the level of prior botanical knowledge and/or detailed field data that is often required.

This work uses an automatic MST (Minimum Spanning Tree)-based reconstruction method to obtain tree skeleton models from LiDAR input data. Multiple scans of the same tree, gathered at different years, are related to each other to improve and expand upon the reconstruction. A procedural model is used to simulate the growth in the tree tips using a lobe-based approach and a region-growing algorithm. The growth-based models provide a temporally informed reconstruction that is visually, geometrically and topologically sound. Establishing correspondences in the main structure between timestamps can assist the reconstruction of the tree at a time for which the input data was noisier or incomplete, as well as provide an estimate of the tree's structure in between known data points. This type of reconstruction can be used to both model and study the growth behaviour of trees, for multi-temporal visualisations, and to provide more informed tree models for reconstruction purposes.

Three-dimensional building models play a pivotal role in shaping the digital twin of our world. With the advance of sensing technologies, unprecedented data acquisition capabilities on capturing the built environment have surfaced, with photogrammetry and light detection and ranging being the two important sources, both of which can acquire point clouds of buildings. A point cloud is anisotropically distributed in space, which---though conveying spatial information itself---has to be converted into a surface model for a wider spectrum of usage. This conversion is often referred to as reconstruction. Despite the enhanced availability of point cloud data in the built environment, how to reconstruct high-quality building surface models remains non-trivial in remote sensing, computer vision, and computer graphics. Most reconstruction methods are dedicated to smooth surfaces represented by dense triangles, irrespective of the piecewise planarity that dominates the geometry of real-world buildings. Although some works claim the possibility of reconstructing piecewise-planar shapes from point clouds, they either struggle to comply with specific geometric constraints, or suffer from serious scalability issues. There is no versatile solution yet for building reconstruction. In this thesis, we propose a novel framework for reconstructing compact, watertight, polygonal building models from point clouds. Our approach comprises three functional blocks: (a) a cell complex is generated via adaptive space partitioning that provides a polyhedral embedding as the candidate set; (b) an implicit field is learnt by a deep neural network that facilitates building occupancy estimation; (c) a Markov random field is formulated for surface extraction via combinatorial optimisation, where an efficient graph-cut solver is applied. We extensively evaluate the proposed method in comparison with state-of-the-art methods in shape reconstruction, surface approximation and geometry simplification. Experimental results reveal that, with our neural-guided strategy, high-quality building models can be obtained with significant advantages over fidelity, compactness and computational efficiency. The method shows robustness to noise and insufficient measurements due to occlusions, and generalise reasonably well from synthetic scans to real-world measurements. Moreover, our method remains generic to not only buildings, but any piecewise-planar objects.

Storing accurate models of complex geometries in a compact way has become an increasingly challenging issue, especially when dealing with large datasets. One of such datasets is Cobra-Groeninzicht's database of all trees in the Netherlands. In the gaming industry, a new technique is being used to generate tree models: the L-system. An L-system stores a string representation of the structural model of a tree, with the added possibility for recursive modelling using growing rules. This format proves a promising alternative to more traditional methods of storing complex geometries. However, it remains unclear whether it can be an accurate enough representation for modelling and analysing real-life trees.

In this research project, the AdTree algorithm is used to reconstruct a skeleton from a point cloud of a single tree. This skeleton is then transformed to an L-System string format, as well as a CityJSON format (both in JSON structure). The L-system format comes with the advantage that it allows for several methods of increasing its compactness further (growing, generalisation). The overall size of these files also indicates fewer storage space is needed to store the tree geometry. The quality of the L-System skeleton is nearly equal to the input, the skeleton generated by. Assuming it can be read and drawn using a Turtle program, the L-system thus allows for storing the same geometric information more compactly than traditional storage formats, with sufficient accuracy, and the added possibilities of growing or generalising the model.

In recent years, 3D building models have become increasingly widespread and are intensively exploited in fields of computer graphics and geometry processing. One of the common surface representations of 3D building models is polygon mesh, which is both compact and efficient in terms of exchange format
and data processing respectively. Downstream applications of polygon meshes can be found in the fields of urban planning, digital mapping, and fluid simulation.
However, the aforementioned applications usually require the input to be watertight and manifold, which is not always fulfilled by existing building models. Moreover, the interior structures of a building model are also considered redundant in certain applications. From such practical demands comes
our graduation project, i.e. trying to recover watertight and manifold outer surface from error-ridden 3D building models.
Existing methods with respect to outer surface extraction can be categorized into two types: surfaceoriented methods and volumetric methods. The former focuses on one particular type of artifacts and operates directly on the surfaces of the defective model. Since surface-oriented methods mainly introduce local operations where needed, unnecessary changes of the original model can be avoided and the result is of high fidelity. Whilst, the latter generates an intermediate representation of the original model, based on which the outer surface is extracted. The volumetric methods are more heuristic for our project since they are designed especially for multiple artifacts and the results are gauranteed with some desired properties.
In this thesis, we propose a hybrid approach for the extraction of outer surface from error-ridden 3D building models, which aims at recovering a watertight and manifold outer shell of the original model. The advantage of our method is that it is non-parametric, fully automatic, and have no assumptions for the input. Moreover, the small features of original model are kept to the greatest extent after processing.
Our method can be divided into four steps: 1) pre-processing, 2) constrained tetrahedralization, 3) classification, and 4) outer surface extraction. All six types of artifacts listed in this paper are gradually resolved during these steps, resulting in a watertight and manifold representation.
The results from our experiments turn out that our methodology can generate valid results in most cases, while preserving input faces and small features at the same time. Comparing with several stateof-the-art methods, our results still possess superior properties in terms of validity and integrity.

3D indoor reconstruction has been an important research area in the field of computer vision and photogrammetry. While the initial techniques developed for this purpose use sensor devices and multiple images for data acquisition and extracting 3D information and representation of the scene, with the advent of deep learning techniques, there has been good progress in extracting 3D information of an indoor scene reconstruction using a single image. This has potential in minimizing user efforts and cost for data acquisition. The current state-of-the-art method involves two main components, the global depth map and plane instances. After investigating the current state-of-the-art methods, it is observed that there is inconsistency in reconstructed surface boundaries and depth estimation over the curvature and edges of the objects present in the scene, despite having good 3D representation in the surrounding regions. We devise a loss function for optimizing depth estimation during supervision of the neural network by providing geometric awareness to the pixels at local level based on its neighborhood properties defined by spatial compatibility and color similarity. A similar function is used during 3D reconstruction for orientation consistency in the point cloud. Based on the quantitative and qualitative analysis, it is observed that the proposed approach helps in improving the 3D reconstruction from a single image in an indoor environment.

Road network maps facilitate a great number of applications in our everyday life. However, their automatic creation is a difficult task, and so far, published methodologies cannot provide reliable solutions. The common and most recent approach is to design a road detection algorithm from remote sensing imagery based on a Convolutional Neural Network, followed by a result refinement post-processing step. In this project I proposed a deep learning model that utilized the Multi-Task Learning technique to improve the performance of the road detection task by incorporating prior knowledge constraints. Multi-Task Learning is a mechanism whose objective is to improve a model's generalization performance by exploiting information retrieved from the training signals of related tasks as an inductive bias, and, as its name suggests, solve multiple tasks simultaneously. Carefully selecting which tasks will be jointly solved favors the preservation of specific properties of the target object, in this case, the road network. My proposed model is a Multi-Task Learning U-Net with a ResNet34 encoder, pre-trained on the ImageNet dataset, that solves for the tasks of Road Detection Learning, Road Orientation Learning, and Road Intersection Learning. Combining the capabilities of the U-Net model, the ResNet encoder and the constrained Multi-Task Learning mechanism, my model achieved better performance both in terms of image segmentation and topology preservation against the baseline single-task solving model. The project was based on the publicly available SpaceNet Roads Dataset.

Deep learning methods have been demonstrated to be promising in semantic segmentation of point clouds. Existing works focus on extracting informative local features based on individual points and their local neighborhood. They lack consideration of the general structures and latent contextual relations of underlying shapes among points. To this end, we design geometric priors to encode contextual relations of underlying shapes between corresponding point pairs. Geometric prior convolution operator is proposed to explicitly incorporate the contextual relations into the computation. Then, GP-net, which contains geometric prior convolution and a backbone network is constructed. Our experiments show that the performance of our backbone network can be improved by up to 6.9 percent in terms of mean Intersection over Union (mIoU) with the help of geometric prior convolution. We also analyze different design options of geometric prior convolution and GP-net. The GP-net has been tested on the Paris and Lille 3D benchmark, and it achieves the state-of-the-art performance of 74.7 % mIoU.

In recent years, there is an ever-increasing demand — both by industry and academia — for 3D spatial information and especially, for 3D building models. One of the ways to acquire such models derives from the combination of massive point clouds with reconstruction techniques, such as Ball Pivoting and Poisson Reconstruction. The result of these techniques is the representation of individual buildings or entire urban scenes in the form of surface meshes — data structures consisting of vertices, edges and faces. Despite their usefulness for visualization purposes, the high complexity of these meshes, along with various geometric and topological flaws, stands as an obstacle to their usage in further applications, such as simulations and urban planning. To address the issue of complexity, the production of lightweight building meshes can be achieved through mesh simplification, which reduces the amount of faces used in the original representations. Moreover, simplification methods focus on conforming the resulting mesh to the original one, in order to minimize the differences between the two of them. As a consequence, simple and accurate building models are possible to be acquired, whose geometric and topological validity is yet questionable. In this thesis, we introduce a novel approach for the simplification of building models, which results into a more compact representation, free of topological defects. The main characteristic of our method is structure awareness — namely, the recovery and preservation, for the input mesh, of both its primitives and the interrelationships between them (their configuration in 3D space). This awareness asserts that the resulting mesh closely follows the original and at the same time, dictates the geometric operations needed for its construction in the first place — thus providing accuracy, along with computational efficiency. Our proposed methodology is divided into three main stages: (a) primitive detection via mesh segmentation, (b) storage of primitive interrelationships in a structure graph and (c) simplification. In particular, simplification is accomplished here by approximating the primitive borders with a building scaffold, out of which a set of candidate faces is defined. The selection of faces from the candidate set to form the simplified mesh is achieved through the formulation of a linear binary programming problem, along with certain hard constraints to ensure that this mesh is both manifold and watertight. Experimentation reveals that our simplification method is able to produce simpler representations for both closed and open building meshes, which highly conform to the initial structure and are ready to be used for spatial analysis. Additionally, a fairly good approximation of a given mesh is possible to be obtained within reasonable execution times, regardless of the initial noise level or topological invalidity. Finally, a comparative analysis shows that the accuracy of our method stands in parallel with that of other available simplification techniques.

Trees are of great significance throughout the world, both in urban scenes and in natural environments. Models of trees can be widely applied in various fields, for instance, landscape design, geo-simulation, environment modelling, and forestry inventories. Recently, laser scanning technology has been rapidly developed, making it possible to effectively acquire geometric attributes of trees and achieve accurate 3-dimensional tree modelling. Existing studies on tree modelling from laser scanning data are vast. Nevertheless, some works don’t ensure sufficient modelling accuracy, while some other works are mainly rule-based and therefore highly depend on user interactions. In this thesis, we propose a novel method to accurately and automatically reconstruct tree branches from laser scanned points. We first employ the Minimum Spanning Tree (MST) algorithm to extract an initial tree skeleton over the single tree point cloud, then simplify the skeleton through iterative removal of redundant components. A global-optimization approach is performed to fit a sequence of cylinders to approximate the geometry of the tree branches. The results show that our approach is adaptable to various trees with different data qualities. We also demonstrate both the topological fidelity and geometrical accuracy of our approach without significant user interactions. The resulted tree models can be further applied in the precise estimation of tree attributes, urban landscape visualization, etc.

The use of 3D models has been rapidly expanding, finding applications in both
scientific and commercial fields. One common requirement for these various applications is the geometrical and topological validity of these models. However, many models available online contain deficiencies in various forms, such as duplicated geometry, gaps in the surface, etc.. To cope with those deficiencies, a standard solution is the clean extraction of the model’s boundary, and simultaneously the model’s reconstruction in a way that its structure is valid. This thesis tackles a more generalized problem, the inside-outside classification for these models. Where many approaches might have requirements for running analysis, the methodology presented strives to robustly handle all cases.
These last decades, there have been various approaches in solving the ”inside -
outside classification problem”. A major attempt utilizes the winding number
algorithm, in order to assign values to elements whose position is relevant to the
input model. By assessing that value, a decision on whether the element in question is interior or exterior is taken. Other approaches work with casting rays, or other geometric analysis to also identify the borders of a model and segment the interior from the exterior. Also, since deficiencies inhibit the kick-starting of the necessary analysis, there are methods that try to restructure said models in order to clear any existing deficiencies. The methodology within this thesis will attempt a different approach from those that have been presented until now, which is transferring the problem from three into two dimensions. The first step is introducing a planar cross section on the area of interest. From there, through some graph reconstruction, geometric and optimization applications, a valid 1-manifold boundary of the cross-section is created. On that, the application of inside-outside classification through ray casting is possible. Assessing the results of the pipeline proves that the automated process can produce valid results, for a particular point of interest, related to an input model. The pipeline has been proven to function regardless of the cutting plane’s orientation, and can handle robustly a multitude of geometrically and topologically defective models. The results from this thesis can inspire further applications, and improvements on the pipeline can further evolve the quality of its outcome.

3D computer models are starting to play a more and more important role in our society. Realworldsituations are often too complex to explain in a 2D map and also the interest in virtualreality, serious gaming and other technologies that can be based on 3D computer models, isgrowing.