Graphics Dissertation Online
Permanent URI for this community
For information about PHD Award please visit Eurographics Annual Award for Best PhD Thesis.
Eurographics PHD Award Winners (under construction)
Browse
Browsing Graphics Dissertation Online by Subject "3D Reconstruction"
Now showing 1 - 3 of 3
Results Per Page
Sort Options
Item 3D Reconstruction and Rendering from High Resolution Light Fields(ETH Zurich, 2015) Kim, ChangilThis thesis presents a complete processing pipeline of densely sampled, high resolution light fields, from acquisition to rendering. The key components of the pipeline include 3D scene reconstruction, geometry-driven sampling analysis, and controllable multiscopic 3D rendering. The thesis first addresses 3D geometry reconstruction from light fields. We show that dense sampling of a scene attained in light fields allows for more robust and accurate depth estimation without resorting to patch matching and costly global optimization processes. Our algorithm estimates the depth for each and every light ray in the light field with great accuracy, and its pixel-wise depth computation results in particularly favorable quality around depth discontinuities. In fact, most operations are kept localized over small portions of the light field, which by itself is crucial to scalability for higher resolution input and also well suited for efficient parallelized implementations. Resulting reconstructions retain fine details of the scene and exhibit precise localization of object boundaries. While it is the key to the success of our reconstruction algorithm, the dense sampling of light fields entails difficulties when it comes to the acquisition and processing of light fields. This raises a question of optimal sampling density required for faithful geometry reconstruction. Existing works focus more on the alias-free rendering of light fields, and geometry-driven analysis has seen much less research effort. We propose an analysis model for determining sampling locations that are optimal in the sense of high quality geometry reconstruction. This is achieved by analyzing the visibility of scene points and the resolvability of depth and estimating the distribution of reliable estimates over potential sampling locations. A light field with accurate depth information enables an entirely new approach to flexible and controllable 3D rendering. We develop a novel algorithm for multiscopic rendering of light fields which provides great controllability over the perceived depth conveyed in the output. The algorithm synthesizes a pair of stereoscopic images directly from light fields and allows us to control stereoscopic and artistic constraints on a per-pixel basis. It computes non-planar 2D cuts over a light field volume that best meet described constraints by minimizing an energy functional. The output images are synthesized by sampling light rays on the cut surfaces. The algorithm generalizes for multiscopic 3D displays by computing multiple cuts. The resulting algorithms are highly relevant to many application scenarios. It can readily be applied to 3D scene reconstruction and object scanning, depth-assisted segmentation, image-based rendering, and stereoscopic content creation and post-processing, and can also be used to improve the quality of light field rendering that requires depth information such as super-resolution and extended depth of field.Item Self-Supervised Shape and Appearance Modeling via Neural Differentiable Graphics(2023) Henzler, PhilippInferring 3D shape and appearance from natural images is a fundamental challenge in computer vision. Despite recent progress using deep learning methods, a key limitation is the availability of annotated training data, as acquisition is often very challenging and expensive, especially at a large scale. This thesis proposes to incorporate physical priors into neural networks that allow for self-supervised learning. As a result, easy-to-access unlabeled data can be used for model training. In particular, novel algorithms in the context of 3D reconstruction and texture/material synthesis are introduced, where only image data is available as supervisory signal. First, a method that learns to reason about 3D shape and appearance solely from unstructured 2D images, achieved via differentiable rendering in an adversarial fashion, is proposed. As shown next, learning from videos significantly improves 3D reconstruction quality. To this end, a novel ray-conditioned warp embedding is proposed that aggregates pixel-wise features from multiple source images. Addressing the challenging task of disentangling shape and appearance, first a method that enables 3D texture synthesis independent of shape or resolution is presented. For this purpose, 3D noise fields of different scales are transformed into stationary textures. The method is able to produce 3D textures, despite only requiring 2D textures for training. Lastly, the surface characteristics of textures under different illumination conditions are modeled in the form of material parameters. Therefore, a self-supervised approach is proposed that has no access to material parameters but only flash images. Similar to the previous method, random noise fields are reshaped to material parameters, which are conditioned to replicate the visual appearance of the input under matching light.Item The Smart Point Cloud: Structuring 3D intelligent point data(ORBi, 2019-06-05) Poux, FlorentDiscrete spatial datasets known as point clouds often lay the groundwork for decision-making applications. E.g., we can use such data as a reference for autonomous cars and robot’s navigation, as a layer for floor-plan’s creation and building’s construction, as a digital asset for environment modelling and incident prediction... Applications are numerous, and potentially increasing if we consider point clouds as digital reality assets. Yet, this expansion faces technical limitations mainly from the lack of semantic information within point ensembles. Connecting knowledge sources is still a very manual and time-consuming process suffering from error-prone human interpretation. This highlights a strong need for domain-related data analysis to create a coherent and structured information. The thesis clearly tries to solve automation problematics in point cloud processing to create intelligent environments, i.e. virtual copies that can be used/integrated in fully autonomous reasoning services. We tackle point cloud questions associated with knowledge extraction – particularly segmentation and classification – structuration, visualisation and interaction with cognitive decision systems. We propose to connect both point cloud properties and formalized knowledge to rapidly extract pertinent information using domain-centered graphs. The dissertation delivers the concept of a Smart Point Cloud (SPC) Infrastructure which serves as an interoperable and modular architecture for a unified processing. It permits an easy integration to existing workflows and a multi-domain specialization through device knowledge, analytic knowledge or domain knowledge. Concepts, algorithms, code and materials are given to replicate findings and extend current applications.