Langguth, Fabian (2018)
Gradient Domain Methods for Image-based Reconstruction and Rendering.
Technische Universität Darmstadt
Ph.D. Thesis, Primary publication
|
Dissertation, Finale Version -
Text
Dissertation_Fabian_Langguth_2018.pdf - Accepted Version Copyright Information: CC BY-NC-SA 4.0 International - Creative Commons, Attribution NonCommercial, ShareAlike. Download (134MB) | Preview |
|
Dissertation Fabian Langguth - Appendix -
Video
Dissertation_Fabian Langguth_Appendix.mp4 - Accepted Version Copyright Information: CC BY-NC-SA 4.0 International - Creative Commons, Attribution NonCommercial, ShareAlike. Download (107MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Gradient Domain Methods for Image-based Reconstruction and Rendering | ||||
Language: | English | ||||
Referees: | Goesele, Prof. Dr. Michael ; Drettakis, Dr. George | ||||
Date: | 18 May 2018 | ||||
Place of Publication: | Darmstadt | ||||
Date of oral examination: | 2 July 2018 | ||||
Abstract: | This thesis describes new approaches in image-based 3D reconstruction and rendering. In contrast to previous work our algorithms focus on image gradients instead of pixel values which allows us to avoid many of the disadvantages traditional techniques have. A single pixel only carries very local information about the image content. A gradient on the other hand reveals information about the magnitude and the direction in which the image content changes. Our techniques use this additional information to adapt dynamically to the image content. Especially in image regions without strong gradients we can employ more suitable reconstruction models and we can render images with less artifacts. Overall we present more accurate and robust results (both 3D models and renderings) compared to previous methods. First, we present a multi-view stereo algorithm that combines traditional stereo reconstruction and shading based reconstruction models in a single optimization scheme. By defining as gradient based trade off our model removes the need for an explicit regularization and can handle shading information without the need for an explicit albedo model. This effectively combines the strength of both reconstruction approaches and cancels out their weaknesses. Our second method is an image-based rendering technique that directly renders gradients instead of pixels. The final image is then generated by integrating over the rendered gradients. We present a detailed description on how gradients can be moved directly in the image during rendering which allows us to create a fast approximation that improves the quality and speed of the integration step. Our method also handles occlusions and compared to traditional approaches we can achieve better results that are especially robust for scenes with reflective or textureless areas. Finally, we also present a new model for image warping. Here we apply different types of regularization constraints based on the gradients in the image. Especially when used for direct real-time rendering this can handle larger distortions compared to traditional methods that use only a single type of regularization. Overall the results of this thesis show how shifting the focus from image pixels to image gradients can improve various aspects of image-based reconstruction and rendering. Some of the most challenging aspects such as textureless areas in rendering and spatially varying albedo in reconstruction are handled implicitly by our formulations which also leads to more effective algorithms. |
||||
Alternative Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-80876 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science 20 Department of Computer Science > Graphics, Capture and Massively Parallel Computing |
||||
Date Deposited: | 07 Dec 2018 09:41 | ||||
Last Modified: | 07 Dec 2018 09:41 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/8087 | ||||
PPN: | 439669324 | ||||
Export: |
View Item |