• Computer Graphic
• Computer Vision
• Image Processing
• Multi-view data

The ever-increasing diffusion of digital cameras and the advancements in computer vision, image processing and storage capabilities have lead, in the latest years, to the wide diffusion of digital image collections.
A set of digital images is usually referred as a multi-view images set when the pictures cover different views of the same physical object or location.

In multi-view datasets, correlations between images are exploited in many different ways to increase our capability to gather enhanced understanding and information on a scene.
For example, a collection can be enhanced leveraging on the camera position and orientation, or with information about the 3D structure of the scene.
The range of applications of multi-view data is really wide, encompassing diverse fields such as image-based reconstruction, image-based localization, navigation of virtual environments, collective photographic retouching, computational photography, object recognition, etc.
For all these reasons, the development of new algorithms to effectively create, process, and visualize this type of data is an active research trend.

The thesis will present four different advancements related to different aspects of the multi-view data processing:

- Image-based 3D reconstruction: we present a pre-processing algorithm, that is a special color-to-gray conversion. This was developed with the aim to improve the accuracy of image-based reconstruction algorithms.
In particular, we show how different dense stereo matching results can be enhanced by application of a domain separation approach that pre-computes a single optimized numerical value for each image location.

- Image-based appearance reconstruction: we present a multi-view processing algorithm, this can enhance the quality of the color transfer from multi-view images to a geo-referenced 3D model of a location of interest.
The proposed approach computes virtual shadows and allows to automatically segment shadowed regions from the input images preventing to use those pixels in subsequent texture synthesis.

- 2D to 3D registration: we present an unsupervised localization and registration system. This system can recognize a site that has been framed in a multi-view data and calibrate it on a pre-existing 3D representation.
The system has a very high accuracy and it can validate the result in a completely unsupervised manner.
The system accuracy is enough to seamlessly view input images correctly super-imposed on the 3D location of interest.

- Visualization: we present PhotoCloud, a real-time client-server system for interactive exploration of high resolution 3D models and up to several thousand photographs aligned over this 3D data.
PhotoCloud supports any 3D models that can be rendered in a depth-coherent way and arbitrary multi-view image collections.
Moreover, it tolerates 2D-to-2D and 2D-to-3D misalignments, and it provides scalable visualization of generic integrated 2D and 3D datasets by exploiting data duality.
A set of effective 3D navigation controls, tightly integrated with innovative thumbnail bars, enhances the user navigation.

These advancements have been developed in tourism and cultural heritage application contexts, but they are not limited to these.